People who rarely produced original writing are suddenly producing extensive design proposals, business plans, documentation, presentations, tickets, pull requests, blog articles, and meeting summaries, all generated by AI. It is all unreadable. AI is helping me write better and faster, but I’m fed up with reading almost anything written by AI. In this essay, I want to express my frustration, and describe how I’m finding value in AI for writing.

AI Is Terrible for Writing

A pattern I see is that people use AI to build something new, then they use AI to retrospectively summarize what they have already built into a design document. Reading a document like this isn’t just difficult—it is punishing. The design document is no longer a proposal to build consensus, bring people along with you, and refine ideas through slow, deliberate thinking—it is a summary made by a machine, in exhausting detail, without context or perspective. It is unreadable. Inhumane. The authors of these documents become impatient when people aren’t engaged, and it is hard not to be, because the thing they built is already working.[1]

I also see more and more pull request summaries clearly written by machines for machines. They are rich in detail: this was changed to that, these things were split, those things were merged, this was left untouched, tests were added for this, and so on. But I’m left asking: Why are we doing this? What is the value? How risky or urgent is this work? Where do you want my input? What should I pay attention to? The same goes for tickets or meeting summaries made by AI. The writing, like more and more of the world, consists of statements outside of context.

I had someone write me a personal message about a sensitive topic that was clearly workshopped with AI in an effort to nuance the conversation and not offend me. But the message became impersonal, dispassionate, and disjointed. It had all of the parts, but it didn’t make sense as a whole. I wasn’t interested in reading it, or responding. I would much rather someone be themselves and write with their own voice, or with some passion, even if they can’t quite find the words, or risk offending me.

In relationships with others, there is vulnerability and risk. In communicating with your AI, there is none, and attempts to eliminate vulnerability and risk will eliminate the process of relationship itself. As Simon Sarris writes in Resist Summary, an essay I keep returning to:

I think one should write as much as they can with their own empiricism, their own senses, giving the reader their own characterization of life or events.
—Simon Sarris, Resist Summary

Perhaps the laziest and most offensive use of AI writing I have seen: I had someone use AI to summarize my comments on their proposal as their response to me! Thanks! I couldn’t have said it better myself![2]

I Don’t Have the Context You Do

Cynthia Dunlop recently shared survey results on reading AI-written articles entitled Report: How developers react to AI-scented blog posts. For the majority of readers, if they think an article is AI-assisted or AI-authored, they will stop reading (78%) and avoid the author in the future (71%). As Bryan Cantrill wrote about this research in his essay The revolt of the reader: “our brains pull an LLM-triggered ejection handle, bailing us out mid-sentence in an act of self-preservation.”

The strongest result in the survey was people overwhelmingly preferred the author’s own writing (98%), with all its flaws and idiosyncrasies, as compared to a soulless rewrite by an AI. Writing authentically is one of the best ways to connect with people. It reminds me of a comment Bjarne Stroustrup, the creator of C++, made during a panel discussion I attended over a decade ago on writing books:

When I read a book, I really feel that if I can hear the author’s voice ringing in my head—complete with weird accents and peculiarities of their speaking patterns—then I feel like something has succeeded. If it feels like dry, academic text, somebody has written yet another dry, academic text. And so, giving talks about the subject really helps. And, for many people, ... I can hear them when I read it, and that is good.
—Bjarne Stroustrup[3]

When you are the one who prompted the AI, the written output is often quite useful. But do not mistake this for effective writing. When you read the output, you already have lots of context—you formulated the prompts, established the constraints, and provided artefacts like documentation, source code, logs, statistics, pictures, etc., that led to the output. You can skim the output and quickly decide what is relevant and valuable versus what is irrelevant or wrong. You are also part of the process—prompting the AI and seeing the results—a call and an answer. You bring a human perspective to the machine. When you send the same text to someone else, they have little to no context. They are not part of the process—there is no unfolding—so it is much harder to decide what is relevant versus what is not. They are forced to read exhaustively and consider every line, peering into the internals of a machine with the hope of establishing context. It is only natural to stop reading.

AI Can Be Great for Writing

I recently wrote an academic paper.[4] I used AI extensively and it made the writing process faster and more enjoyable. But AI did not write a single line of the paper. I wrote the paper in LaTeX. The context I provided to the AI included the journal style guide and LaTeX template, all papers previously published in the journal, the source code for the software systems I was writing about, and the configuration, logs, and metrics for production deployments of the software.

After writing a paragraph, many times I would ask the AI to ensure what I had written was correct by referencing the source code or configuration, or examining production logs or metrics. For example, ensuring I had correctly described which columns in the database were indexed, or how the rows in the Parquet files were sorted. This allowed me to stay in the context of writing, often moving on to the next paragraph while the AI verified the previous one. It helped me spot important omissions and inaccuracies. Remarkably, reversing this process—asking the AI to write the paragraph from the context I provided, like describing how something works from the source code—was never valuable. Not once. It was consistently unpleasant to read, and often inaccurate.

Another thing that helped me stay in context was using AI to complete the citations. It is tedious to look up the details of a paper, website, or book, fill in the BibTeX format, and then reference it in the text. Instead, I would leave parenthetic instructions for the AI to reference what I wanted, and then keep on writing without breaking stride.[5]

The AI was incredibly good—indeed ruthless—at finding spelling and grammar mistakes. It was great at suggesting simplifications to sentences that are too long or unclear.[6] It was skilled at drawing technical diagrams in TikZ, again speeding up the work, and saving me from learning this syntax. The AI even found a very subtle mistake where I had used the incorrect notation in one paragraph, something four human reviewers who were experts in the system all missed.[7]

What was the one thing the AI was incredibly good at writing? It wrote the abstract—the most terse, mechanical, inhuman part of the whole paper—indeed, the most abstracted part. I included it entirely unchanged. Perhaps this shouldn’t be a surprise, as AI is great at summary, abstraction, and re-presentation of original work.

Can AI Become Good at Writing?

It would be a mistake to underestimate the writing that can one day be achieved by AI. We are still in the very early days of model development, and the models most of us are using are trained to be factual, direct, and correct, rather than creative. However, Murat Demirbas’s excellent essay The Safest Job from AI may be Writing notes that the quality of writing from LLMs has plateaued and it may be difficult to improve since unlike programming, mathematics, and accounting, it is difficult to provide a clear formulation and verifiable output.[8]

Since LLMs lack an active mental model of a specific human reader, they are just optimizing for the statistical probability of the next word over a vast dataset. They cannot empathize with the human reader, as they don’t have the human lived experience. And they have zero skin in the game.
—Murat Demirbas, The Safest Job from AI may be Writing

I am interested in two efforts to improve AI writing that I’ve recently come across. The first is ASD-STE100 Simplified Technical English, a standard for writing technical documentation that grew out of an aerospace working group in the 1980s. It has two parts: a controlled dictionary and writing rules about grammar and style. A version of ASD-STE100 has been created for use with AI models.[9] ASD-STE100 isn’t appropriate for all AI-generated writing—it is specifically targeted at clear and easy-to-follow instruction manuals. I’m going to give it a try for installation instructions, runbooks, and similar documentation.

The second is Pangram, a model fine-tuned to detect AI writing. Bryan Cantrill in the essay The revolt of the reader notes:

To use an LLM to write is to void the social contract between writer and reader: we readers shouldn’t be expected to labor to understand a sentence that the writer themselves didn’t work to create.
—Bryan Cantrill, The revolt of the reader

I could not agree more. Bryan has started experimenting with Pangram to detect AI-generated writing that violates this contract and has gone so far as to mandate that all public writing at Oxide be reported by Pangram as human-authored. A tool like this could save us from the laziest and worst AI writing.[10]

When There Are No Words

When we write, the meaning of our words is not always direct, factual, or concise. Examples include poetry, literature, song, mythology, humour, and theatre. Iain McGilchrist, who I believe is one of the most important writers of our generation, writes in the book The Matter with Things: “The more important something is, the more we have to struggle in the attempt to reduce it to language.” Ask an LLM what the reliable formula is for falling in love. In AI-generated writing, detailed information gives the illusion of meaning. But in mimicking human knowledge, intelligence, and understanding, an AI writing about the facts is very different from the quality of writing from a lived experience. The whole cannot be decomposed into parts without a substantial loss of meaning.

The direct approach destroys its object. Similarly, there are things that can be conveyed only indirectly. To recapture the depth of what is said, or in order to say anything truly new, we are always engaged in going beyond everyday language, which means being alert to what is being said implicitly.
—Iain McGilchrist, The Matter with Things

It might seem silly to say that writing in business, engineering, or science should involve uncertainty, not just facts and conclusions. But operations, incident response, product development, performance engineering, hunting for security vulnerabilities, performance reviews, mission statements, and many other endeavours, are all creative and involve uncertainty. The narrative is regularly as important as the facts and conclusions. It is also easy to anthropomorphize LLMs as good systems thinkers because they can synthesize such a wide range of information. However, returning to Simon Sarris: “narrative resists compression” and “the opposite of summary is attention to detail”.

Despite an emphasis on linear, first-principles thinking, the work of a good systems engineer, just like the work of a good anthropologist or psychologist, is difficult to put into words, and especially into summary. You can’t necessarily observe and write things down using a formula. Even writing things down loses some of the context. The observer is also part of that context. Systems work always has some sense of discovery, holding open the unknown, an unfolding of relationship with the system itself, and a willingness to be surprised. Normal language doesn’t always have the words.[11]

The nature of the output, of summary, is speed things along. But the consequence is to avoid you having to build your own complex mental model of anything. I worry that without a complex model in one’s own mind, one may never notice complex relationships that are otherwise missed. A loss of attention to detail.
—Simon Sarris, Resist Summary

Conclusion

I don’t want to live in a world where you use AI to summarize something important into unreadable text, and then I use AI in an attempt to decipher it. I want to hear you, imperfections and all. I want your interpretation of aesthetics, beauty, quality, relationship, time. I want to know how you feel. I want you to cut through and tell me what really matters. I don’t want to read a summary of facts and connections separate from experience and relationship. As Iain McGilchrist describes, the way we attend to the world changes the nature of what comes into being, the only world we can know.[12]

If you ask a machine to do something for you, you are at risk of not asking for what you want, but what it can do well. If what it can do well is summarize, and you rely upon the tool, you may tune your questions so that you get better answers. This looks like training a machine to do what you want, but its also training yourself to ask questions about the world in a certain way. What are the questions you are omitting?
—Simon Sarris, Resist Summary

Intentional writing will likely become more valuable. People who write, and write to think, to think deeply and carefully, or to create, to share, or to capture something important without explicitly expressing it will continue to write and produce original work. The people who never were writers will use AI to produce lots of text.

As we read more and more text generated by AI, we might not notice what AI has disrupted and changed, because we’ve become more shallow anyway. The search for certainty limits our humanity. Sometimes it is more important to not find the words. I don’t want to read what you didn’t write.

1. In some instances, design documents are not the right fit, and using AI to rapidly collaborate around working software is perhaps a better approach to refine ideas and build consensus. ↩︎

2. In another example, a group of people had been working on a proposal, building consensus through a number of wiki pages. Someone new wanted to modify the proposal. Rather than going through the messy, relational process of working with this group to update the existing wiki pages, they used AI to proliferate a new wiki page, copying the existing work and adding their own twist. ↩︎

3. I’ve used this quote previously in Techniques for Improving Your Technical Writing. ↩︎

4. The paper describes a database I created and my hope is it will be accepted to The Conference on Innovative Data Systems Research (CIDR). ↩︎

5. For example, I would write: {TODO: reference the Amazon Aurora paper} and the AI would take care of the rest. ↩︎

6. I also use AI as an editor for my blog. It gives me more confidence posting an essay without having someone else proofread it. But I only let AI identify errors or make suggestions. I don’t let it rewrite and change my voice or perspective. ↩︎

7. I was describing a log-structured merge (LSM). We had originally called the levels Generation 1, Generation 2, Generation 3, or Gen1, Gen2, Gen3 for short. In the paper, I described the LSM more generically using L0, L1, L2 for Level 0, Level 1, Level 2, respectively. This meant my colleagues reading this paper had to map L1 to Gen2 in their heads, which made my error in one paragraph of incorrectly using L1 when I meant L2 extra hard to spot. The LLM found it, no problem. ↩︎

8. Another interesting essay on the limitations of AI for writing, especially the individuality in understanding, is The Human Skill That Eludes AI: Why can’t language models write well? by Jasmine Sun in The Atlantic. ↩︎

9. Thanks to Chris Riccomini for making me aware of this. ↩︎

10. For more on Pangram, listen to the Oxide and Friends podcast with the founder Max Spero. One thing they note is there is some evidence that people who use LLMs regularly are good at detecting LLM writing, but people who do not are less adept. There is also an interesting discussion about trying to reverse the prompt and the input context in an effort to describe how much of the output came from the input and how much came from the model. ↩︎

11. The best systems thinkers I know tend to be excellent writers. I expect this skill to remain important and perhaps become even more important. Returning to Murat Demirbas’s essay on writing that I referenced earlier, it concludes: “And maybe programming itself is turning into a form of creative writing, getting more opinionated and more architectural.” ↩︎

12. Iain also draws the distinction between comprehension—taking the world as a whole—and apprehension—taking hold of the world. The quality of attention is different. I find it interesting that while apprehension is the act of seizing or capturing, it is also a fearful or uneasy anticipation of the future. I highly recommend watching Iain’s lecture AI and the Battle for the Soul: Information is Not Understanding at Ralston College. ↩︎

Way to cheapen the experience, Apple

(Image credit: Shutterstock / fizkes)

• iOS users are seeing ‘ads’ for other Apple products on their iPhones
• The banners promote services like iCloud+, Apple Music and AppleCare+
• The messages often can’t be dismissed and stick around for weeks

If you were to ask iOS users how they felt about Apple’s decision to insert ads into Apple Maps, it would quickly become obvious that it was not a popular move. So when iPhone fans discovered that Apple has been inserting prompts for its own products into iOS, the reaction was predictably swift and furious.

The ‘ads’ in question appear in the Settings app in iOS. They can be found near the top of the iPhone’s display and promote several products to the user, including paid-for iCloud+ storage, free trials for services like Apple Music and Apple TV, and the AppleCare+ extended warranty.

Infuriatingly for iOS users, the only way to disable these promos is to either let them expire — which in many cases could take weeks or even months, leaving a persistent notification badge on the Settings app in the meantime — or pay for the service being offered. There is not always a “dismiss” button to shoo them away for good, and when there is, some people have reported that it doesn’t seem to work.

Many people have taken to social media to vent their frustrations with Apple’s actions. “I’m feeling like this is nearly on-par with the ens***ification that Microsoft introduced with Start menu ads,” said one aggravated user. “I wish Apple would just stop that crap,” stated another. Many other people have taken to Reddit to express their exasperation, suggesting that the discontent is widespread.

Some of these ads and offers — such as the one for Apple Music — appear after you’ve bought a new Apple product. With the company launching the iPhone Duo and iPhone 18 Pro range recently, it’s perhaps unsurprising that we’re seeing a sudden influx of posts flush with irritation and indignation.

Apple’s services push

(Image credit: Future)

On the face of it, Apple’s decision to include promotions like this in its operating system might seem like a strange one. After all, Apple has long positioned itself as a premium company where the user experience is paramount. From that viewpoint, repeatedly attempting to prod customers towards even more of the company’s products might seem cheap, even distasteful. And when you’ve just shelled out a record fee for an iPhone after Apple’s painful price rises, that can just feel like nickel and diming.

But Apple has been putting increasing emphasis on its services in recent years, of which products like iCloud+, Apple Music, Apple TV and AppleCare+ are all a part. With limited room for growth in the smartphone market, the company is turning to areas where it thinks it can make more progress.

And unfortunately for users, that means more ads like this on their devices. While they are far from the most obnoxious ads around, the fact that many cannot be dismissed is clearly rubbing users the wrong way — as is the fact that the promos even exist at all in an Apple product.

There is also the possibility that at least some of these adverts are appearing due to a bug. The iCloud+ one in particular has drawn comments from users who are already subscribed to iCloud+ and thus, one would think, should not need to see any ads like this. Time will tell how that plays out and whether Apple fixes this ‘bug’ or leaves the ads in place. If the latter transpires, it would suggest that they are not appearing erroneously.

With Apple’s services push showing no signs of slowing down and further ads supposedly planned for the likes of the firm’s Visual Intelligence system, don’t expect Apple to have a change of heart any time soon.

Watermarks that spy on users are no mere watermarks

There’s a sneaky new evolution of the “watermark”, let’s call it — the spymark.

A watermark is a visible mark embedded in a physical or digital medium to verify authenticity or assert ownership.

A spymark is a hidden signal that makes your work traceable without your knowledge or consent.

Spymarking is sneaking onto the internet

Google SynthID is a spymark that embeds secret hidden signals “imperceptible to humans” (Google’s own words) into images, audio, text, and video. This signal can encode database identifiers that map to your identity. Your user records, full name, IP addresses, date of birth, physical addresses, political party affiliation, and more.

Google’s SynthID-Image paper reports that its SynthID-O variant can encode a 136-bit payload in a 512x512-pixel image. That is enough room for a 64-bit database identifier, with 72 bits left for error correction.

SynthID was not the first spymark system designed, and it’s hardly the only one under active development. OpenAI and many other tech companies are developing these systems at scale. These companies claim spymarking can help identify AI-generated content, yet they’ve gone beyond simple watermarking and built in robust tracking mechanisms. Social media, content production tools, and smartphones may soon find themselves filled with spymarking algorithms that sneak these signals into everything you publish.

For example, images can be invisibly altered in their frequency domain to carry tracking information such as database IDs linked to users:

A smaller version of Mochi’s photo contains a real toy watermark: ID 173, encoded through subtle pixel changes. Compare the original and marked image, animate the amplified differences, and decode the ID from the PNG. The associated author and timestamp are fictional.

Why give it a new name?

“Watermark” has become a catchall for historical marks of authenticity, banknote security features, copyright overlays, and hidden tracking signals in our media. That last usage obscures the privacy risk that some forms of modern “watermarks” have become.

Most people tire of the discourse on privacy. It’s complex, repetitive, and seems irrelevant to the typical day-to-day routine of the average citizen. We can’t keep explaining the technicalities behind statistical tracking tools each and every time we need to communicate these concepts to people and expect them to pay attention.

“To speak the name is to control the thing.” — Ursula K. Le Guin, The Rule of Names

With a simple change of terminology we can put the privacy concern up front, cementing this issue in the conversation forever:

• Spy- — clandestine surveillance, involuntary disclosure
• -mark — embedded signal

Now we get to disambiguate friendly watermarks from concerning new technology while leveraging the familiar etymology. It collapses a technical communication salient and gains us a lot of ground with just one new word. It’s clear, coherent, and straight to the point.

No more wasting energy establishing the basic facts. The privacy concern is built into the word.

“Spymark” is a privacy Rumpelstiltskin.

More spymark examples

Here are a few more forms of spymarks so you can better familiarize yourself.

Audio spymarks operate using principles similar to image spymarks, and they are typically inaudible. Some methods make minute changes to the audio waveform in the time domain; others modify features in the frequency domain, and some combine both approaches. These methods can encode data, including identifiers linked to personal information:

Compare spectrograms of the same LJ Speech excerpt unwatermarked and with Timbre, AudioSeal, WavMark, FSVC, Patchwork, or Norm-Space. The authors’ original plots are linked below.

Spectrograms from Wen et al. (2025), SoK: How Robust is Audio Watermarking in Generative AI models? · Original samples. Plots retain the authors’ original scales.

These schemes are engineered to be robust. They can often survive compression or re-encoding.

These tools are already proliferating, and in fact they predated generative AI and the push to watermark generative media. The open source spymarking tool audiowmark originated in 2018 and can hide 128-bit payloads in audio and protect them with a secret AES key, preventing users without the key from decoding them.

You can even encode personal information invisibly into text! SynthID steers word choices to create a detectable statistical pattern that can encode a tracking payload:

An illustrative encoding: eight word choices represent eight bits. The binary value 10101101 gives database ID 173, which can point to a record containing an author and timestamp. These are fictional examples, not a SynthID decoder.

Watermarks are benign, spymarks are not

Watermarks remain easy for the user to spot and are typically not nefarious.

Sometimes watermarks deter counterfeiting:

Sometimes watermarks claim ownership:

Dorothea Lange, Migrant Mother (1936) — Getty sells some public domain images.

And sometimes they’re just annoying:

But these watermarks don’t track you.

Printer tracking dots, on the other hand, are an early example of spymarks, dating back to the 1980s:

Image: EFF — Robert Lee, Seth Schoen, Patrick Murphy, Joel Alwen, and Andrew “bunnie” Huang

Standardized, user-editable tags are not spymarks

We need to be clear that spymarks are a form of metadata you have limited knowledge of and control over, and that their purpose is entirely antagonistic to you.

Metadata such as the EXIF tags in photos and the ID3 tags in MP3 files are standardized, well-documented fields. While EXIF can expose sensitive data such as GPS coordinates, these standardized fields can be inspected, edited, and removed from files you control.

Inspect editable EXIF and ID3 metadata in two tabs. EXIF shows fictional camera-maker, model, and author tags in a JPEG metadata segment; ID3 shows an MP3’s title, artist, and album. Hover or tap the bytes for explanations. These samples contain metadata only.

You can strip standard metadata tags out of your files. Unfortunately, a spymark signal embedded in pixels, audio, or word choices is invisible to you and can remain in your files even after you edit them.

Furthermore, whereas tags are helpful for maintaining information such as song titles or a photo’s exposure settings, spymarks encode user-tracking identifiers that are entirely opaque and useless to you. These signals can survive metadata removal and some edits, allowing marked copies to remain traceable as they circulate.

Keeping our files free of spying

It’s hard to imagine the future where we can’t even trust our own files. And the sad thing is that this has already started.

Spymarks are certainly not great for whistleblowers or anyone who doesn’t want to be persecuted for their words or affiliations. No matter where you stand on whatever issues, spymarks can be used against you and those you care about.

Imagine a future where every device is attested and every social media post carries an account-linked spymark. Where every file or post served to you for sharing has spymarks embedded within. Combined with account records and observations of where copies appear, these identifiers can help reconstruct how content spreads. In this world, it would be dangerously easy to hunt down anyone from a JPEG or a tweet. Or track the web of humans through which “dangerous ideas” flow.

That future lies halfway between now and 1984. So let’s stay off that timeline, shall we?

Call a spymark what it is. A spy tool used to spy on you and everyone you interact with.

Usage

spy·mark /ˈspaɪmɑːrk/ noun

plural spymarks

A hidden signal embedded in media to trace its origin, tools, or distribution history without the user’s meaningful control.

“The spymarks tied each copy to a different recipient.”

spymark verb, transitive

spymarked; spymarking; spymarks

To embed a spymark in a file or piece of media.

“The social network spymarks every uploaded image.”

spymarking noun

The practice of embedding spymarks.

“The platform introduced spymarking without telling its users.”

spymarked adjective

Containing a spymark.

“Removing the file’s metadata still left the image spymarked.”

FoxDev Studio opens the projects, forms and tables you already have and runs them the way you remember: no rewrite, no conversion, no export step.

Download Get started How it is built

It picks up where you left off

Point it at a folder you have not opened in years, and what is there is what you get.

Your project, as it is

Projects, forms, class libraries, menus and reports open straight from the files you already have. Nothing is migrated first, and nothing is left behind.

The designers you know

Forms, classes, menus and reports are edited where you expect them to be, alongside the project manager, the Command Window and a debugger that stops on your line.

Your data stays your data

Tables, indexes, memos and databases are read and written in place. What sits on disk afterwards is the same kind of file it was before.

Whatever it talked to, it still talks to

The system calls, the automation objects and the old add-in libraries your application leans on keep working, so the parts nobody wants to touch stay untouched.

What it looks like

The IDE, on Visual FoxPro's own sample projects. Every picture is a real session.

Visual FoxPro's own Solution sample, imported and running: its launcher form on the Screen tab, the project's 123 forms, 7 menus and 11 class libraries in the explorer, and the runtime waiting for events.

The editor lints through the runtime's own compiler, so what it underlines is what would fail.

Stopped one step past a breakpoint in the Solution sample's main program: call stack, locals, watches, and the line it is on.

A database container in the table browser: 8 fields, 17 records, memos a click away.

FoxScript: lambdas answering HTTP routes, a 32-bit library called from inside one, listening on 8080.

Built to match, not to approximate

Small differences are what break an old application: a number printed a column too wide, an event arriving a moment late, an error with the wrong number on it. So behaviour here is settled by asking Visual FoxPro itself and matching its answer, rather than by reading a reference page and hoping.

What comes out of that is a runtime written from scratch: quick to start, self-contained, and straightforward about the corners it has not reached yet.

1,722 elements of the Visual FoxPro 9 language reference known to the runtime

1,534 of them exercised by a test that compares the answer with the product's

3 names it has not met yet; everything else runs, is ignored on purpose, or is refused by name

What Visual FoxPro 10 would have been

Visual FoxPro stopped at version 9, and at 32 bits. This is the same language, rebuilt on a foundation that has not been frozen since 2007. Four pieces are what make that possible.

The ceiling that came with 32 bits is gone

Visual FoxPro is a 32-bit program, and that decides more than it appears to. It is why a table stops at two gigabytes, why a memo file stops at two gigabytes, and why a big report runs out of memory on a machine with plenty to spare. The limits are signed 32-bit numbers buried in the file handling, not a licensing decision anybody made.

FoxDev Studio is 64-bit throughout. Every file offset is 64-bit and a table is never read into memory at all, so the same .dbf that used to stop dead carries on into the hundreds of gigabytes. One thing to know before you lean on it: a table grown past two gigabytes will not open in Visual FoxPro again. If you still work in both, that is a one-way door.

A compiler, and a machine built to run what it makes

Visual FoxPro compiled your program to p-code and shipped a runtime to execute it. This is the same arrangement, made again: a compiler and a bytecode interpreter written in Rust and compiled to WebAssembly, so one machine runs your code wherever the application runs. The editor checks what you type through that very compiler, so what it underlines and what the runtime refuses cannot drift apart.

A running program is a fiber. When it needs something from the world outside (a message box, a modal form, the next record) it does not call out and block; it yields, the work is done while the machine is off the stack, and the answer is handed back. That is why MESSAGEBOX() stops your program without freezing the window behind it, why READ EVENTS waits without spinning, and why SetFocus can fire GotFocus, and Init can run while a form is still being built, in the order FoxPro always did it.

How the machine works and what it runs.

The form you see is the object tree, not a picture of it

A running form is a live tree of objects with the properties you would expect, and the interface is drawn straight from that tree by React. Each object watches only itself, so THISFORM.lblGreeting.Caption = cMsg repaints one label rather than the whole form. On a dense screen that is the difference between instant and sluggish.

The same tree is what the designer edits, one step earlier. There is no second model kept in step with the first, which is the usual place a form and its designer start telling different stories.

Your old libraries still load, though nothing here is 32-bit

An .fll is a 32-bit image, and every process in a 64-bit application is 64-bit, so nothing inside the application itself could ever open one. Rather than tell you it is impossible, SET LIBRARY TO starts a small 32-bit process whose only job is to hold your library, and the runtime talks to it. The calls are synchronous, because a program may call into a library halfway through an expression, and an answer that arrived later would not be an answer. It is measured against real libraries: the encryption library, FoxTools, and libraries built from Microsoft's own API samples.

Nothing 64-bit needs the bridge: DECLARE ... DLL reaches a modern library in the same process, and automation objects are reached the way they always were. The old road stays open; it just is not the only one any more.

FoxScript: the same language, with more room

Everything you have written still means what it always meant. FoxScript only adds on top: a block you can hand to something else to run later, and a way to answer a web request from the code that already knows your business.

server.prg / FoxScript

&& your old add-in libraries load just as before
SET LIBRARY TO "vfpencryption71.fll" ADDITIVE

LOCAL oServer
oServer = FoxScript.Http.CreateServer()

oServer.Get("/api/v1/customers/:id", LAMBDA(req, res)
    LOCAL lnId
    lnId = VAL(req.Params("id"))

    SELECT * FROM customer WHERE cust_id = lnId INTO CURSOR c_cust

    IF RECCOUNT("c_cust") > 0
        res.Status(200).Json(FoxScript.Data.CursorToJson("c_cust"))
    ELSE
        res.Status(404).Json('{"error": "Not found"}')
    ENDIF

    USE IN c_cust
ENDLAMBDA)

oServer.Listen(8080)
READ EVENTS

Every line of that runs on the same runtime your forms do. The queries, the cursor and the library call are ordinary FoxPro; the lambda and the server are what FoxScript adds: no second language, no service to stand up beside it. The keywords and the HTTP API are each written down in full.

Download

The nightly is rebuilt from every push to main and published as a pre-release on GitHub. Unsigned, so the first launch asks you to confirm.

Windows The installer (.exe) and the runtime (.zip), x64.

Ubuntu and other Linux An AppImage and a .deb, x64.

Source, on GitHub FoxDevCommunity/FoxDevStudio: the IDE, the VM, this site, and every release.

Read the documentation

Each part of the product is written down, including the parts that are not there yet.

Overview What FoxDev Studio is, and where each part of it is described.

Getting started Install the IDE, open a project, run a form, ship an executable.

Required tooling What you need to run the IDE, and what you need to build it from source.

The virtual machine Fibers, host requests, and why the VM never blocks.

Bytecode The module format, how a program compiles, and what a frame holds.

Instruction reference Every instruction with its operands and stack effect, generated from the VM source.

The new keywords LAMBDA, ENDLAMBDA, the FoxScript namespace and two new value types.

The HTTP API FoxScript.Http, the Node side of it, and why it is shaped this way.

What is coming

Work already mapped out, roughly in the order it is being built.

The last of the language A short list of corners is still open, each one checked against the original as it closes.

Reports The report designer and the report engine, band by band, measured the same way the rest was.

Tables past two gigabytes, as a format of their own The 64-bit offsets are there today; a container that is honest about the DBF header's limits comes next.

An installer for what you ship Your application handed to the people who use it as a single thing they can run, with nothing to set up first.

FoxDev Studio is an independent project, and one still growing. The documentation says where it is and where it goes next.

Model summary

Intelligence

#1 / 212

58

Artificial Analysis Intelligence Index

4 out of 4 units for Intelligence.

Speed

N/A

Output tokens per second

Unknown out of 4 units for Speed.

Cost

#93 / 212

In $4.00 Out $20.00 Cache Discount 95%

$5.98

Cost per Intelligence Index task

4 out of 4 units for Cost.

Verbosity

#95 / 212

260M

Output tokens from Intelligence Index

4 out of 4 units for Verbosity.

Comparison Summary

Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) is amongst the leading models in intelligence, but somewhat expensive when comparing to other models of similar price. The model supports text and image input, outputs text, and has a 1M tokens context window.

Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) scores 58 on the Artificial Analysis Intelligence Index, placing it well above average among comparable models (median: 25). When evaluating the Intelligence Index, it generated 260M tokens, which is very verbose in comparison to the median of 88M.

Pricing for Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) is $4.00 per 1M input tokens (somewhat expensive, median: $2.00) and $20.00 per 1M output tokens (somewhat expensive, median: $10.00). On average, it costs $5.98 per task to evaluate Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) on the Intelligence Index.

Technical specifications

Reasoning

Yes

This page shows the reasoning version of this model.

A non-reasoning variant may also exist.

Input modality

Supports: text and image

Output modality

Supports: text

Context window

1M

~1500 A4 pages of size 12 Arial font

212 models in this class

Metrics are compared against models of the same class:

• Non-reasoning models → compared only with other non-reasoning models
• Reasoning models → compared across both reasoning and non-reasoning
• Open weights models → compared only with other open weights models of the same size class:
  • Tiny: ≤4B parameters
  • Small: 4B–40B parameters
  • Medium: 40B–150B parameters
  • Large: >150B parameters
• Proprietary models → compared across proprietary and open weights models of the same price range, using a blended 3:1 input/output price ratio:
  • <$0.15 per 1M tokens
  • $0.15–$1 per 1M tokens

  • $1 per 1M tokens

Intelligence

Artificial Analysis Intelligence Index

Artificial Analysis Intelligence Index v4.3.2 incorporates 10 evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, AA-LCR v1.1

Artificial Analysis Intelligence Index v4.3.2 includes: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, AA-LCR v1.1. See Intelligence Index methodology for further details, including a breakdown of each evaluation and how we run them.

Open Weights

Indicates whether the model weights are available. Models are labelled as 'Commercial Use Restricted' if commercial use is limited by conditions, and as 'Non-commercial' if the license prohibits commercial use.

Capability Indexes

Measures the performance of models on specific capabilities and industries.

Artificial Analysis Finance & Accounting Index

Incorporates 7 evaluations: AA-Omniscience, GDPval-AA v2.1, AA-Briefcase v1.1, Humanity's Last Exam, AutomationBench-AA, AA-LCR v1.1, GDP.pdf · Higher is better

Benchmarks

Intelligence Evaluations

Intelligence evaluations measured independently by Artificial Analysis · Higher is better

• AA-Briefcase v1.1 — Agentic knowledge work, (Elo-500)/2000
• GDPval-AA v2.1 — Agentic real-world work tasks, (Elo-500)/2000
• AutomationBench-AA — Agentic SaaS workflows
• Terminal-Bench 4.0 — Agentic coding & terminal use
• SciCode — Coding
• Humanity's Last Exam — Reasoning & knowledge
• GDP.pdf — Professional document reasoning, All-pass
• CritPt — Physics reasoning
• AA-Omniscience Accuracy — Knowledge
• AA-Omniscience Non-Hallucination Rate — 1 - hallucination rate
• AA-LCR v1.1 — Long context reasoning
• Harvey LAB-AA — Legal agentic work, criterion pass rate
• EnterpriseOps-Gym-AA — Agentic business operations
• AA-AnalystAgent — Quantitative analysis on spreadsheets & documents
• 𝜏³-Banking — Agentic tool use
• ITBench-AA — Kubernetes incident root-cause analysis
• MMMU-Pro — Visual reasoning
• MLCR-AA — Medical long context reasoning

Intelligence Evaluation Relevance

While model intelligence generally translates across use cases, specific evaluations may be more relevant for certain use cases.

Artificial Analysis Intelligence Index

Artificial Analysis Intelligence Index v4.3.2 includes: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, AA-LCR v1.1. See Intelligence Index methodology for further details, including a breakdown of each evaluation and how we run them.

AA-Briefcase v1.1

AA-Briefcase v1.1 is an agentic knowledge work benchmark developed by Artificial Analysis. AA-Briefcase Elo is a combined metric that aggregates rubric pass rate, analytical quality Elo and presentation Elo · Higher is better

AA-Briefcase Elo is a combined metric that aggregates analytical quality Elo, presentation Elo, and rubric pass rate, with rubric performance converted into Elo via synthetic head-to-head matches. Elo and 95% confidence interval bounds are clamped at 0.

AA-Omniscience

AA-Omniscience Index (higher is better) measures knowledge reliability and hallucination. It rewards correct answers, penalizes hallucinations, and has no penalty for refusing to answer. Scores range from -100 to 100, where 0 means as many correct as incorrect answers, and negative scores mean more incorrect than correct.

Intelligence Index Comparisons

Cost per Intelligence Index Task

Weighted average cost per Intelligence Index task. Each evaluation’s cost is calculated from input, cache hit, cache write, reasoning, and answer token prices, divided by task count, and weighted by its Intelligence Index weight.

Artificial Analysis Intelligence Index

Artificial Analysis Intelligence Index v4.3.2 includes: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, AA-LCR v1.1. See Intelligence Index methodology for further details, including a breakdown of each evaluation and how we run them.

Token Use

Output Tokens per Intelligence Index Task

Weighted average number of output tokens used to run one task in the Artificial Analysis Intelligence Index.

The number of tokens required per Intelligence Index task. This is calculated by multiplying the output tokens per eval by the relative weights of each benchmark in the Intelligence Index, then dividing by task count (excluding repeats).

Cost

Cost per Intelligence Index Task

Weighted average cost per Intelligence Index task, segmented by token type. Lower is better.

Cost to Run Artificial Analysis Intelligence Index

The cost to run the evaluations in the Artificial Analysis Intelligence Index, calculated using the model's input, cache hit, cache write, reasoning, and answer token prices and the number of tokens used across evaluations (excluding repeats).

Cache Hit

Price per token for cached prompts (previously processed), typically offering a significant discount compared to regular input price, represented as USD per million tokens. The values shown here are the cache hit price; cache write and cache storage are billed separately and vary by provider — see "Cache pricing by provider" for detail.

Context Window

Context Window for RAG

Larger context windows are relevant to RAG (Retrieval Augmented Generation) LLM workflows which typically involve reasoning and information retrieval of large amounts of data.

Context Window

Maximum number of combined input & output tokens. Output tokens commonly have a significantly lower limit (varied by model).

Frequently Asked Questions

When was Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) released?

Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) was released on September 22, 2026.

Who created Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback)?

Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) was created by Anthropic.

How intelligent is Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback)?

Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) scores 58 on the Artificial Analysis Intelligence Index, placing it well above average among other reasoning models in a similar price tier (median: 25).

How much does Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) cost?

Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) costs $4.00 per 1M input tokens (somewhat higher than average, median: $2.00) and $20.00 per 1M output tokens (somewhat higher than average, median: $10.00), based on Anthropic's API.

What is Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) API pricing?

Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) costs $4.00 per 1M input tokens and $20.00 per 1M output tokens (based on Anthropic's API). For a blended rate (7:2:1 cache hit/input/output ratio), this is $2.94 per 1M tokens. Pricing may vary by provider. Compare provider pricing

How verbose is Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback)?

When evaluated on the Intelligence Index, Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) generated 260M output tokens, which is at the higher end compared to other reasoning models in a similar price tier (median: 88M).

Is Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) a reasoning model?

Yes, Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) is a reasoning model. It uses extended thinking or chain-of-thought reasoning to work through complex problems before providing an answer.

What input modalities does Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) support?

Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) supports text and image input.

What output modalities does Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) support?

Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) supports text output.

Can Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) process images?

Yes, Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) supports image input and can analyze, describe, and answer questions about images.

Is Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) multimodal?

Yes, Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) is multimodal. It can process text and image input and generate text output.

What is the context window of Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback)?

Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) has a context window of 1.0M tokens. This determines how much text and conversation history the model can process in a single request.

Is Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) open source?

No, Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) is proprietary. The model weights are not publicly available.

How many parameters does Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) have?

Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) is a proprietary model and Anthropic has not disclosed the model size or parameter count.

How does Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) perform on benchmarks?

Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) achieves a score of 58 on the Artificial Analysis Intelligence Index. This composite benchmark evaluates models across reasoning, knowledge, mathematics, and coding.

Is Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) available via API?

Yes, Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) is available via API through 6 providers. Compare API providers

Where can I use Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback)?

Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) is available through 6 API providers. Compare providers

California solar panels tower over canals in Turlock, proving to save water, produce clean energy and improve local water quality.

Solar panels cover part of a Turlock Irrigation District canal as part of Project Nexus in Hickman, California, on Sept. 4, 2026. The pilot project is testing the use of solar panels over working irrigation canals to generate renewable energy. (Beth LaBerge/KQED)

This story is part of Flipping the Switch, KQED’s series about California’s transition to clean energy and what it means for you. What works? What doesn’t? And how much does it cost?

Driving from Los Angeles to San Francisco, Jordan Harris saw the concrete canals that transport water across the state slowly evaporating under the scorching West Coast sun.

Harris said he spent a lot of his life in France, where he noticed how trees lined the canals and protected their precious water from the wind and sun. “And so I started thinking, could we do something similar in California?”

California’s extensive network of canals and aqueducts is the lifeblood of the state, providing clean water to populous cities and expansive farms that wouldn’t otherwise have it. While the state has long struggled with evaporation, researchers at UC Davis estimate that California’s total water supply could shrink by up to 25% by 2050, the equivalent of two Lake Shastas. They attribute the decline, in part, to hotter temperatures caused by climate change.

In an effort to limit warming poised to dry out the state, California pledged to shift to 100% clean electricity by 2045.

Solar panels cover part of a Turlock Irrigation District canal as part of Project Nexus in Hickman, California, on Sept. 4, 2026. (Beth LaBerge/KQED)

The state currently generates 62% of its electricity from renewable and zero-carbon sources, with the largest percentage of that coming from solar. But meeting California’s energy demand will require an unprecedented expansion of solar and storage capacity.

Thinking back to the baking concrete canals, Harris pitched an “obvious” solution to his friend Robin Raj: Why not build solar panels on top of California’s 4,000 miles of open canals that transport water across the state? The panels would produce clean energy while limiting evaporation by shading the canal — a win-win.

The duo went on to start Solar Aquagrid, an advocacy firm dedicated to reimagining aging canals, and partnered with the Turlock Irrigation District, UC Merced and the California Department of Water Resources to cover small sections of Turlock’s canals with solar panels. The new “solar canal” provides shade to limit evaporation and generates power all in one neat package. They call it Project Nexus, and it operates at the intersection of the state’s need for clean water and energy.

Project Nexus tests not one, but three different solar canal models. A wide-span canopy rises over one of Turlock’s 115-foot canals, resembling the top of a barn. A more classic set of slanting solar panels sits atop a 20-foot canal. And accordion-like retractable panels are positioned low across the narrow canal, ready to roll up at a moment’s notice.

Raj affectionately called the wide-span canopy his “singular field of dream.”

Roger Bales, a professor at UC Merced who co-authored a 2021 study that helped convince the state to invest in this pilot, said his initial calculations indicate that putting panels over 100 miles of Turlock’s narrow canals could save enough water for 2,700 households annually.

If the state put panels over 100 miles of the wide canals, it could save 11,000 households’ worth.

He also estimates that the narrow and the wide canal could generate 330 to 1,400 megawatts per hundred miles respectively — enough to power 77,000 to 304,000 households.

The pilot was fully completed and energized last year. Earlier this month, Harris walked alongside the solar panels that arched over a 115-foot canal.

Solar AquaGrid co-founders Jordan Harris (left) and Robin Raj walk alongside solar panels covering part of a Turlock Irrigation District canal in Hickman, California, on Sept. 4, 2026. Solar AquaGrid helped develop Project Nexus, a pilot project testing the use of solar panels over working irrigation canals. (Beth LaBerge/KQED)

“These are disturbed spaces,” Harris said, pointing to the concrete ditch filled with murky water. “So why not take advantage of the fact that we’ve already done that and help them work harder, be more efficient, produce more and protect what’s around us?”

The research team also found numerous “co-benefits,” or silver linings that make the project more valuable than the sum of its water and power.

Bales said shading the canals led to less algal growth and less fixed aquatic weed growth — reducing maintenance and improving water quality.

Preventing new solar development on natural land benefits both people and planet, Bale added, emphasizing how preserving these lands is essential for maintaining life “as we know it on the Earth.”

Acquiring permits to build on undeveloped land can be a major bottleneck for developers, too. Building on open space takes 621 days on average, compared with 110 days for projects on developed land, according to a study from The Nature Conservancy.

A Turlock Irrigation District canal in Hickman, California, on Sept. 4, 2026. (Beth LaBerge/KQED)

Given the numerous benefits, Bales said the project “just makes sense.”

But just because something makes sense does not mean it’s easy to implement. “We often talk about the challenges of getting to the obvious,” Harris said.

Projects touching land, water and power often involve coordinating many stakeholders. California’s canals can be managed by state or federal water agencies, as well as municipal governments and local irrigation districts. These operators often do not own the land around the canal, let alone the infrastructure to transmit the power to utilities — one of the most expensive parts of renewable energy projects, according to Harris.

Harris and Raj found their “ideal” partner in Turlock Irrigation District, a 139-year-old irrigation district operating 250 miles of canals that provides both water and power to nearly 5,000 farmers in the Central Valley growing everything from beans to peaches.

It is one of four irrigation districts in California providing both water and retail power, noted Josh Weimer, a spokesperson for Turlock.

Weimer said that the project received little pushback from locals, in part because, while the Irrigation District used some staff time for the project, a $20 million state grant paid for both the research and the infrastructure.

Weimer described the project as a “proof of concept” that generates just 1.7 megawatts of energy, a tiny fraction of the 56 megawatts that they generate with a solar farm in the desert.

But the canal site hosts more than just panels. Bales said they “have instruments out there to measure temperature, relative humidity, wind speed, incoming radiation, outgoing radiation, and a prototype instrument to measure evaporation directly.”

When it comes to algae, the panels helped. “Last year when we drained the canal at the end of the irrigation season, there was a physical line on the sun side of the canal that had algae growth and then not algae growth from where the shade was,” Weimer said.

Turlock’s solar panels tower above the water because the irrigation district cleans the canals on a regular basis with equipment.

Solar panels cover part of a Turlock Irrigation District canal as part of Project Nexus in Hickman, California, on Sept. 4, 2026. (Beth LaBerge/KQED)

“This isn’t gonna be the only way to do it,” Weimer said — describing how they drag a massive anchor chain through the water to rip out weeds. Similar panels in India and Arizona sit low over the canals, leaving little room for the kind of cleaning that Turlock needs to do.

Over in Contra Costa County, the urban landscape makes any kind of solar canal project unrealistic, said Jennifer Allen, a spokesperson for the county’s water district.

The Contra Costa Water District manages 48 miles of canals winding from Knightsen to Martinez. The Bureau of Reclamation built the system back in the 1930s when the region was predominantly farmland — much like Turlock.

Today, the canal system runs through backyards and public parks.

Something like Project Nexus wouldn’t work because maintenance workers need to access the canals and have limited space to work alongside them, Allen said.

Solar AquaGrid co-founder Robin Raj stands near solar panels covering part of a Turlock Irrigation District canal in Hickman, California, on Sept. 4, 2026. The company helped develop Project Nexus, a pilot project testing solar panels over working irrigation canals. (Beth LaBerge/KQED)

Even though the canal is fenced in, “we have instances where we have cars go into the canal — we’ll have people go into the canal, and animals go into the canal,” Allen said, describing some of the challenges they faced as an urban water district.

“What Turlock is doing with their projects is very impressive, but we have a different-looking canal system going through our service area,” Allen said. They are looking to put their canals into pipes to avoid evaporation, instead.

Andrew Schwarz, a spokesperson for the state’s Department of Water Resources, said that they are waiting to see the final report before deciding to invest in more projects. “These canals vary widely — from narrow, unlined dirt channels that deliver irrigation water to individual fields, to concrete-lined channels over a hundred feet wide that transport water across California,” Schwarz said.

The state remains focused on “evaluation, continued study, potential scaling and lessons learned.”

“Critical infrastructure change is really hard, and if you don’t have science behind it to justify the expense and be able to quantify the benefits, then you’re not gonna scale,” Harris said.

“We are fortunate to live in California where this is a priority and where our state agencies are open to this level of examination and exploration,” Raj said. “That can’t be taken for granted.”

The final report comes out later this year.

We’re sharing Unreal Agent — an agent harness that delivers up to 40% cost savings compared to Codex on production workloads and coding/science benchmarks, without any negative performance impact.

If you’re interested in frontier cost-efficiency for your AI agents, we’d love to work together! Get in touch: contact@unreallabs.ai.

While deploying agents in the wild, we wanted them to respond to users quickly and be cost-effective to run. We’ve noticed that agents spend a lot of time and tokens managing tool calls, which motivated us to build Unreal Agent with a harness that would reduce the model’s tool-management overhead.

The Unreal Agent harness manages tool calls in a completely asynchronous way, relieving the underlying model of the need to manage waits, polls, and heartbeats for tools.

This approach drives two major benefits. First, it always allows users to steer the agent without the need to wait for tool calls to finish. Second, it allows the agent to schedule more useful tool call work between model calls, driving frontier cost efficiency. The current version achieves up to 40% cost savings compared to Codex and up to 20% compared to Pi in real workloads and on agentic benchmarks, which we share here.

We believe harness design is a research area in its own right, with many promising ideas still to be researched and implemented.1

Motivation and Architecture

If you try to build an agent-first product, you’ll quickly realize that there’s no golden path for implementing one. Big-brand vendors offer different SDKs to build agents, each with a different set of trade-offs that might not be immediately apparent.

At Unreal Labs, we have built a number of agentic products and learned a few things about popular SDKs along the way.

For example, CLI-oriented SDKs such as Claude’s Agent SDK carry assumptions about local sessions, subprocesses, and resource limits that don’t translate neatly into production use. Handling completion, cancellation, and background tasks reliably often means building your own lifecycle management around them.

Supporting other providers adds compatibility work: switching API modes can break tools or compaction, while SDK upgrades can change message formats and force integration rewrites. Heavy dependency trees add maintenance and supply-chain risk to a runtime we already need to understand and patch ourselves.

Security and approvals that rely on harness hooks and specialized tools, in our experience, tend to require more maintenance and be less robust than deterministic environment or sandbox constraints, outside the harness: allowed/disallowed hosts, granular access tokens, proxies with approval gates.

Along with these technical motivations, we also wanted to build a harness that could always accept user steering messages without delay and juggle heterogeneous tool calls without extra cognitive load for the model. For example, we wanted the agent to be able to kick off a dev environment setup that might take minutes, while exploring the codebase and searching the web in parallel, all without extra token tax.

Every time Unreal Agent issues a tool call, we immediately append an event-log record that the tool has returned in the “in-progress” state, while continuing its execution in the background. Once a tool actually finishes, we append the result into the session log and call an LLM. Making this work without breaking cache was an interesting engineering challenge in itself.2

Watch the animation.

Cost Efficiency

On the surface, Unreal Agent achieves the same outcomes with fewer model turns and fewer input tokens.

We attribute cost savings to two factors:

1. Minimal harness footprint and careful engineering of tool output usage. Unreal Agent has simple prompts, token-optimized tool results, and no sub-agents or workflows.3
2. More tool work per model turn. Unreal Agent has a straightforward asynchronous tool-calling model that is clearly explained to an LLM. This allows it to issue more heavy tool calls per model turn without wasting tokens on polling or waiting for them.

Benchmarks

We’ve built Unreal Agent to deliver real production workflows for us, but it looks good in the benchmarks too. We tested it with GPT-6 Astra xhigh and compared it with Codex and Pi. Here are some of the results.

There are marginal differences in pass rate, which we attribute to benchmark variance.

Terminal-Bench 4.0

GPT-6 Astra · xhigh. Codex (lb) is the leaderboard baseline; Unreal Agent and Pi runs are linked below.

Agent Rate Total $ In/trial Out/trial Turns Tools Harbor unreal-agent 57.9% 1428 1.73M 32k 28 37 27133053 Codex (leaderboard) 57.9% 2350 — — — — — Pi 55.0% 1827 2.83M 35k 44 57 6ccd097a

SWE-Atlas Codebase QnA

Agent Rate Total $ In/trial Out/trial Turns Tools Harbor unreal-agent 65.8% 936 898k 15k 16 27 3d2fa057 Codex 63.3% 1303 1.69M 17k 22 21 11a440fb Pi 64.0% 1033 1.29M 16k 24 60 da4ac972

DeepSWE 1.1

Agent Rate Total $ In/trial Out/trial Turns Tools Harbor unreal-agent 72.4% 1367 1.60M 28k 26 38 2311ca63 Codex 69.0% 1633 2.19M 30k 30 29 e20ecafd Pi 69.6% 1584 2.21M 30k 40 75 cd7d8de6

Agents’ Last Exam · ALE-CLI

Full pass rates and mean scores are listed below. These runs are not on Harbor.

Agent Full pass Mean score Total $ In/task Out/task Turns Tools unreal-agent 30.0% 59.7 217 0.76M 18k 18 23 Codex 29.0% 58.1 292 1.59M 15k — 21 Pi 29.0% 59.2 262 1.19M 19k 27 37

We run mostly coding benchmarks because they are available on Harbor, which makes reproduction and verification easier, but the harness is domain-agnostic.

Getting started with Unreal Agent

The Unreal Agent SDK currently offers:

• Go library that you can integrate directly into your codebase
• Runner executable similar to claude -p / codex exec
• Benchmark runner compatible with Harbor

Check out our github repo if you want to try it for yourself

References

[1] Melissa Z. Pan, Shuo Yang, Negar Arabzadeh, Wei-Lin Chiang, Ion Stoica, and Matei Zaharia. “HarnessTax: How Much Does Harness Matter for Coding Agents?” 2026. ↩

[2] The use of two tool-call result items (one in progress, one final) in a single context is underspecified in the Responses API documentation. During testing, we encountered rejections with some models on some inference providers (not OpenAI), and the function_call_output status field seemed to have no impact in those cases.

Our tests showed that models can understand the progression from a running update to a final result when the conversation format is accepted. We believe this pattern should be explicitly supported by the Responses API and consistently supported across inference providers. ↩

[3] Diogo. “(KV) Cache Rules Everything Around Me.” Complete Skeptic, September 9, 2026. ↩