Improving 15 LLMs at Coding in One Afternoon. Only the Harness Changed (opens in new tab)

I really enjoyed this article. I think the author is precisely right and I've been saying this for a long time. There's a ton of extremely interesting low hanging fruit that can vastly improve the effectiveness of even currently existing models hiding in how we design our agent harnesses; enough to — at least until we hit diminishing returns — make as much or more of a difference than training new models!

I think one of the things that this confirms, for me at least, is that it's better to think of "the AI" as not just the LLM itself, but the whole cybernetic system of feedback loops joining the LLM and its harness. Because, if the harness can make as much if not more of a difference, when improved, as improvements to the model itself, then they have to be really considered equally important. Not to mention the fact that models are specifically reinforcement learned to use harnesses and harnesses are adapted to the needs of models in general or specific models. So they necessarily sort of develop together in a feedback loop. And then in practice, as they operate, it is a deeply intertwined feedback loop where the entity that actually performs the useful work, and which you interact with, is really the complete system of the two together.

I think thinking like this could not only unlock quantitative performance improvements like the ones discussed in this blog post, but also help us conceive of the generative AI project as actually a project of neurosymbolic AI, even if the most capital intensive and a novel aspect is a neural network; and once we begin to think like that, that unlocks a lot of new options and more holistic thinking and might increase research in the harness area.

Seems like a very cool technique, but also very oversold. He's seeing a 5% improvement on a find and replace benchmark of his own devising and saying stuff like this in the blog post:

> Here is why that is backwards. I just showed that a different edit format improves their own models by 5 to 14 points while cutting output tokens by ~20%. That’s not a threat. It’s free R&D.

He makes it sounds like he got a 5-14% boost on a top level benchmark, not 5% improvement on a narrow find and replace metric. Anecdotally, I don't usually have a lot of issues with editing in Claude Code or Cursor, and if there is an issue the model corrects it.

Assuming that it costs double the tokens when it has to correct itself, and find and replace errors are as prominent in actual day to day use as his benchmark, we're talking a 5% efficiency gain in editing token use (not reasoning or tool use). Given that editing must be less than 1/3 of the token use (I assume much less?), we're talking an overall efficiency gain of less than 1%.

This seems like a promising technique but maybe not a high priority in efficiency gains for these tools. The messianic tone, like assuming that Google cut off his access to suppress his genius editing technique rather than just because he was hammering their API also leaves a bad taste, along with the rampant and blatant ChatGPTisms in the blog post.

Great post. A few choice quotes:

> Often the model isn’t flaky at understanding the task. It’s flaky at expressing itself. You’re blaming the pilot for the landing gear.

> The model is the moat. The harness is the bridge. Burning bridges just means fewer people bother to cross. Treating harnesses as solved, or even inconsequential, is very short-sighted.

> The gap between “cool demo” and “reliable tool” isn’t model magic. It’s careful, rather boring, empirical engineering at the tool boundary.

> Codex uses apply_patch: It takes a string as input, which is essentially an OpenAI-flavored diff, and instead of relying on a structured schema, the harness just expects this blob to follow a strict set of rules. Since OpenAI folks are without a doubt smart, I’m sure the token selection process is biased to fit this structure at the LLM gateway for the Codex variants of GPT, similar to how other constraints like JSON schemas or required tool calls work.

Codex does in fact use a schema for constrained sampling, it's here: https://github.com/openai/codex/blob/main/codex-rs/core/src/...

It still has to work to get an exact match, or at least I didn't read the code to see if there's any fuzzy matching used.

Note the two codex models were the only ones doing worse with the author's proposed format. The author found them doing better with replace than with apply patch, but since the author appears to be unaware that they use a schema for constrained sampling, I think a more realistic benchmark should enable constrained sampling for the apply test.

This makes sense to me because I've been having very accurate results with models from even 2+ years ago... but I had to "hold them right." Even when reasoning models and coding agents were just a gleam in Altman's and Amodei's eyes, I could tell a lot of the unrealized gains lay in building the right tools, harnesses and guardrails to manage the context and guide the model. (Relevant subthread as example: https://news.ycombinator.com/item?id=44171519)

But this article hints at deeper wins to be had. Consider that these models are operating on source code, which is a verbose, noisy, textual serialization of the intended syntax / semantic trees. TFA improves accuracy by retro-fitting some structure onto the text. But what if models could operate directly on these underlying structures themselves?

As a data point, there are projects like OpenRewrite, which encode a ton of information, from formatting to types with globally resolved dependencies for each symbol in what they call a "Lossless Semantic Tree", so that there is ~0 ambiguity about the code. When I worked with OpenRewrite (in the era before LLMs, how quaint!) compared to other tools, it produced the best results for code transformations with the highest fidelity to the surrounding code.

Now imagine if the agent has access to such detailed information. It would not have to waste tokens figuring incidental things out like formatting. Although I haven't tested it out myself, I believe Moderne (the maintainers of OpenRewrite) when they say that agents armed with LST-based tools make extremely accurate changes.

This is essentially the same reason why the answer to "Which is better, Vim or Emacs?" is "IntelliJ."

Now consider that these models are STILL operating on text as an input and output mode! What if they were multi-modally trained on source code and docs and their syntax / semantic trees? I don't even know what this would look like, but I'd bet this would produce the most accurate coding models ever -- probably neurosymbolic in the truest sense.

The harness matters far more than most people think. This post about the CORE benchmark where Opus’ score almost doubled when they switched to Claude Code from their own harness. https://x.com/sayashk/status/1996334941832089732

I implemented this hash (read and edit) approach in tilth if you want to test it out.

https://github.com/jahala/tilth

its on npm and cargo:

- cargo install tilth

- npx tilth

then tilth install claude-code/windsurf/cursor --edit

(--edit flag is needed)

I made "tilth" a few days ago, since I'm consistently trying to get the LLMs to use tools more efficiently and spend less tokens doing it -- original tilth post from Monday: https://news.ycombinator.com/item?id=46952321

During my first LLM experiments in Emacs using gptel, I also found that the LLM has considerable difficulties changing source code files with the Unix patch tool.

As Emacs has a built-in tree-sitter package, I implemented this same idea. I created gptel tools like tree_sitter_list_nodes, tree_sitter_get_nodes, tree_sitter_update_nodes, tree_sitter_insert_before_node and tree_sitter_insert_after_node. The "list" tool returns a list of AST nodes with first line number, first line content and node hash. The LLM can then use "get" to collect interesting nodes in their entirety and "update" to update a list of nodes identified by hash with new content (var/function bodies).

Worked like a charm.

Shows how much room for improvement there is on the harness level.

Agents waste a lot of tokens on editing, sandboxes, passing info back and forth from tool calls and subagents.

Love the pragmatic mix of content based addressing + line numbers. Beautiful.

My personal notes (not the author): have been way faster performance wise which is honestly the biggest improvement over correctless. I've posted https://github.com/can1357/oh-my-pi before, but didn't seem to gain traction. It's a great little agent.

What was the point of Claude code or Gemini banning the OP? Why would they care about how IDEs use the underlying API?

I think this is the right take. I usually am aligned with most of what Anthropic is doing, but cutting off OAuth login from open harnesses was a bad move. My guess is there is some serious worry/overlap with the Cursor's of the world - e.g. folks who will be competitors in the future who are taking advantage of cheaper Opus rates/loss leader from them while simultaneously building a competitive model (Composer).

Also, nice clever optimization here. Lots of low hanging fruit in harness land.

It’s funny to see where we are on model improvements.

Back when I was maintaining a coding harness around the time of Claude 3.5 we tried hash prefixes we tried line number prefixes we tried a lot of different approaches to making the model better at selecting edit blocks and ultimately at-least then fuzzy string matching won out.

My experience as well. People worry our profession is being reduced to "prompt engineer", but actually I get the feeling that programming will soon be mainly about designing and building harnesses for specific tasks.

I independently invented a very similar method and then abandoned it because it relies on abstraction.

Instead I now use Damerau-Levenshtein distance to assert the edits to be replaced and if the similarity is over some threshold the edit goes through

Really works well because it's explicit. Forcing the model to emit the source tokens to be replaced seems to improve things.

https://github.com/day50-dev/sidechat/blob/db9c8f9d834967442...

It will often chomp white space differently but the main problem is

1. Track alignment with the lines being tracks (hash fixes that)

2. Content alignment with the model not losing focus (hamming/levenshtein other similarity scores fixes that)

If we demand exact matches we're simply not going to get them.

(Combining both methods might be good, I hadn't thought of that)

Another crucial point: the error line "Content mismatch. Reread the file" is crucial. Errors should give descriptive remediate actions.

So even with crappy models it does this automatically and will tool loop accordingly.

Asking it to do smaller edits is no good. Many smaller models will go down to single line edits, looking around for blank lines and just inject garbage. So don't suggest it.

Larger models, which succeed in doing this, know to do that. Smaller models which don't, won't do it if you don't suggest it

Seriously this thing works with 4B models

I also combine it with a toolcall hack for models that don't support tool calling

https://github.com/day50-dev/sidechat/blob/db9c8f9d834967442...

It injects the tool description in the system prompt after probing the capabilities and then does a simple response router.

I haven't found a model within reason that this doesn't work with (I'm sure if you intentionally throw some fine tune botch up that's emitting garbage it'll break - that's not the claim)

YMMV, works for me™

You forgot to mention your tool does worse for 8/16 LLMs compared to replace?

Problem is, replace has been around for so long, most LLMs are tuned for it now

Great article. To me, this highlights a key question in the era of rapidly advancing machine intelligence: if we know machine intelligence is progressing, what is more valuable to build for? As humans, we still find many tools useful even when doing knowledge work. For instance, a calculator. Sure, a smart person can perform calculations in their head, but it’s much easier to teach everyone how to use a calculator, which is 100% reliable in its intended domain.

In this era, we should build these kinds of tools for problems we know are straightforward ones you can’t get smarter than, even as intelligence continues to advance. Using tools like "bash" or command-line interfaces originally designed for humans is a good initial approach, since we can essentially reuse much of what was built for human use. Later, we can optimize specifically for machines, either accounting for their different cognitive structures (e.g., the ability to memorize extremely long contexts compared to humans) or adapting to the stream-based input/output patterns of current autoregressive token generators.

Eventually, I believe machine intelligence will build their own tools based on these foundations, likely a similar kind of milestone to when humans first began using tools.

I’d really like to see this optimized for the 50-120B parameter open source models that are local viable (gpt-oss-120b, qwen3-80b-3a etc.).

For them I think it would be optimal to provide a tag per function and trust the llm to rewrite the function. As the article notes full reproduction is generally more reliable than edited for short code.

The token and attention overhead from a per line hash I suspect limits this approach for smaller models

There is so much work we can do with harnesses that can make the already existing models so much more capable. I definitely feel the author's frustration as I've also been working on some harness stuff. When Anthropic subscriptions got cut off from OpenCode and other third party tools, I was very disappointed because the model I do the most work in is Claude and I was specifically developing a change [1] in the hopes it would make Claude even better. After that, I started implementing the feature in Claude Code directly (using tweakcc) and after a day of working on that, they even block my tweaked Claude Code with the same message. It means I simply won't be able to use this idea with Claude at all

[1]: the README.md describes the Context Bonsai features in my fork here: https://github.com/Vibecodelicious/opencode

Intriguing, but I wonder if they've looked at whole-conversation token usage, though, and not just short tasks.

I just saw a [paper](https://arxiv.org/pdf/2602.05447) that investigated similar aspects of TOON (which aims to reduce JSON tokens), and they found that even though TOON itself reduced the number of tokens, LLMs were less familiar with it, and thus spent even more tokens trying to decipher it, or making mistakes (see section 4.5, figures 6 and 7).

From the paper: >Unlike Markdown, where each grep hit simply returned more text, TOON's overhead was driven by a combination of output density and additional tool calls from pattern unfamiliarity

---- There's a strangeness tax with LLMs, and it can be substantial.

I would not be surprised at all if this technique turned out to be only a local minimum, with detrimental global effects.

Getting banned from Gemini while attempting to improve Gemini is the most Googley thing ever :D imagine letting your automated "trust and safety" systems run amok so that they ban the top 0.01% of your users with no recourse. Google really knows how score an own-goal.

We build MCP servers that wrap fund APIs. The biggest performance variable we’ve found isn’t the model, it’s how much domain context the harness provides before the model has to reason. Same model, generic prompt versus one loaded with our procedural docs - wider gap than switching between model generations. Which surprised me.

The post’s framing is right but undersells what the harness actually does in production. It’s your trust layer: what can the model touch, what can’t it, how cheaply do you recover when it gets something wrong. We spend something like 70% of engineering time on the recovery path, not the inference. Whether that ratio is right I’m not sure, but it’s where we’ve ended up.

On MCP overhead downthread: real, yes. In regulated environments you need the audit trail and the kill switch, and a tool boundary is how you get those. The unsolved part is keeping the protocol thin enough that you’re not burning tokens on ceremony.

Yea, LLMs have prompt-, harness- and even random seed variability, and it leaves you wonder maybe with a better prompt or system instruction a model could perform better. Too bad most benchmarks don't report that variability, because it could reveal that the model may only perform well if it's prompted in the style of their training data and not generalize well to unseen prompt styles. Also it could explain some of the benchmark vs real world usage gaps.

I remember some papers about earlier models having around 15% prompt variability, and with different tool use sometimes there are even more significant jumps. And if I remember correctly the reasoning models improve some of these because lot of the early prompting tricks is included in them like "thinking step-by-step", "think carefully" and some other "magic" methods. Also another trick is to ask the models to rephrase the prompt with their own words because that may produce prompt that better align with their training prompts. For sure the big model developers are aware of these and constantly improving it, I just don't see too much discussion or numbers about it.

Harness is where the open source should shine. It doesn't require millions of dollars of compute but the search space is vast and explorable with limited budgets.

We know Anthropic likes vertical integration and their walled garden. OpenAI seems to be okay with custom clients using their flat rate subscriptions. But what about Google? Would be great to have a second model that allows use by any client of their flat rate subscriptions. Agentic work with APIs seems to be insanely expensive, so clients need to be able to use the flat rate subscriptions unless you have big bucks.

Witness the giant leap forward in the capabilities of coding agents over the last year. There has been no such leap in LLM model performance. I think the causality is crystal clear. It's nothing about "AGI" and all about existing LLMs learning to use existing tools.

Even a sub-par LLM, put into a context where it has access to unix tools and network and files etc, is vastly more capable than the best LLM chatbot.

The harness is the model "body", it's weight the cognition. Like in nature they develop together and the iteration of natural selection works at both.

If smaller labs (Zai, Moonshot, deepseek, mistral..) get together and embrace a harness, like opencode for example, as a consortium just by the power of "evolution across different environments" they might hit jackpot earlier than bigger labs.

On first principles it would seem that the "harness" is a myth. Surely a model like Opus 4.6/Codex 5.3 which can reason about complex functions and data flows across many files would trip up over top level function signatures it needs to call?

I see a lot of evidence to the contrary though. Anyone know what the underlying issue here is?

Look at Serena. Does something on this line.

> it can use code-centred tools like find_symbol, find_referencing_symbols and insert_after_symbol.

https://oraios.github.io/serena/01-about/000_intro.html

I feel a lot of confusion at which coding harness is best and what options to use. tbh I have mostly used standard aider and I don't know what the consensus is on this tool.

I feel I want to write my own and that maybe in the future a lot of developers will have custom harnesses and have highly customized versions as each user of these models wants to use these things in a way that's unique to their brain, much like how emacs is so great for the customization but one persons emacs config is often not what another wants or only wants a subset and then write their own features.

As an aside what is the feeling on all the various ai coding tools, does aider suck is that aider-ce/cecli are better or are the bespoke tools for each model like claudeCode and such better.

Putting it out there: if any frontier model provider starts allowing any agent to use their $20/month plan, we will all switch to you. We don't want to be forced into 1 harness, we want OAuth, and we want respectable limits without excessive budgets.

Great work, but concurrency is lost.

With search-replace you could work on separate part of a file independently with the LLM. Not to mention with each edit all lines below are shifted so you now need to provide LLM with the whole content.

Have you tested followup edits on the same files?

So the new implementation always operates at the line level, replacing one or more lines. That's not ideal for some refactorings like rename where search and replace is faster.

Edit

Checking ohmypi The model has access to str replace too so this is just a edit till

I made a benchmark on the top tier models as this is not the case in the article. Also I did several cases and the result is self speaking. Hashline is not an improvement that speaks for itself. It is the overhead of harness itself in tool descriptions and schemas. Here are my benchmark results and also my repo for a plugin to truely reduce token usage in top tier models: https://github.com/ASidorenkoCode/openslimedit

I wonder if we'll get to "VI for LLMs" - if the model was trained on using that kind of text navigation and you show context around cursor when it navigates.

Would also be worth having special tokens for this kind of navigation.

I ran into this from the other direction. I built a small SRE agent for my cloud infra and just kind of walked into hand-rolling some of the tools rather than using what exists today. I provided an edit_file tool that felt like it was of reasonable capability, but in practice the agent was regularly 'trying' to do a one line change and submitting PRs that hallucinated 3/4s of the file.

Seeing how bad the results are when you're casually approaching something makes it very evident that it's a topic that can be optimized.

My guess always was that - if you took the source of training data- meaning the authors of the "best" answers and solutions on stackoverflow or github- and got the question reformatted, to sound like it was created by these experts- the created code, would try to hug these sources of truth while getting created.

So, the challenge is actually to find a map of "problem" to "author" and then from "author" to "related code" and from their to a solution.

Seeing all these 'coding' benchmarks reminds me that people still don't understand what coding means in practice. People still think one-phase puzzle-solving is coding. Real coding almost always has multiple phases which build on top of one another. There is an architectural component which is missed here - and the sheer number of phases/layers is actually where most of the complexity comes from.

This matches my experience exactly. I’ve been building an MCP server with 82 tools and spent weeks on infrastructure testing. Switching from a Docker-based Cloudflare Tunnel to a native tunnel process took my tool call success rate from ~50% to 100% — same model, same tools, same prompts. The harness isn’t just important, it’s often the dominant variable.

One of the first things I add to my claude instructions file is to stop using grep, its awfully slow, just use ripgrep instead, you can just type the word of what you're looking for from the project root and find it all in one shot. Claude likes to go folder by folder with grep and it drives me crazy.

"You're absolutely right!"

At this point I'd take a contract with Anthropic to have Claude code pick better tooling.

> Treating harnesses as solved, or even inconsequential, is very short-sighted

Is it possible that burning extra tokens is the point, since they get paid more?

Underrated is how much improving harnesses, not just models, has a lot to do with productive uses of LLMs at tasks like coding in the last year.

Has any harness matched the effectiveness of Claude Code yet? I haven't experimented much recently, but every time I have in the past, I wasn't able to get any other tool to approach how effective I am in CC.

I'd love to use a different harness-- ideally an OSS one-- and hook it up to whichever LLM provides the best bang for the buck rather than being tied to Claude.

I feel like cursors solution is still the best answer. Let the model suggest edits in whatever format it prefers using as few "extra" tokens as possible and have a small model figure it out. I don't use cursor anymore but when I did it was impressive how consistently it worked, I think there was a single time it failed. 70b might be overkill though...

I do agree with his identification of the problem: sometimes agents fail because of the tools around it and not because of the model's reasoning. However, for the failing tests I think he is not making the distinction between a failed test due to a harness failure or due to a reasoning failure. It would be nice if someone analyzed that from the data set.

Yep this has been my experience with browser agents as well. One little change in the harness/agentic loop and the model suddenly becomes a whole lot smarter at navigating the web. I was also able to build a better browser agent than ‘claude —chrome’ in just a few afternoons just by tweaking the harness.

When you're in the business of selling tokens - you look at technology that reduces that as a threat. If they were selling services that USE tokens, then reducing them would be welcome... so they'll likely steal this and incorporate it into their proprietary CLIs like claude code...

It seems like agentic (or atleast AI-assisted coding) is the future. And we will be increasingly relying on these models to earn our livelihood.

Is anyone else worried at how easily Anthropic/Google/OpenAI can basically cut you off if you do something they don't like?

I feel the baseline comparison should be relative to the intuitive and simple "line-numbers only" schema.

It's less token heavy than the proposed hash approach, and I don't think frontier LLMs hallucinate line numbers if each line in the context is prefixed with them.

You can improve a lot the success rate with providing HELM and clear instructions with the tool description.

Over a year ago had a lot of issues and the description and example was the difference between 30-50% failure to 1%!

So I'm surprised a bit about the point. May be I'm missing it.

Yeah I invented a similar method for information extraction attribution around 2022, I would place custom markers in a document so the extraction model can reference them together with the answer and be unique on the document to be able to locate it.

I enjoyed the post. Sometimes small hacks go a long way. I still feel like the game changer will be models outputting just enough code for replacement and harness replacing it efficiently rather than whole code with find and replace hacks.

Non-native speaker here. Can someone please be so nice to explain why do we use the word "Harness" here and not e.g. Orchestrate or Steer?

It took me some time to realise what people mean by it, originally confusing it with harvest.

Interesting. I wonder if that's one of the reasons Claude Code works so incredibly well for me with Clojure — I use clojure-mcp tools, which provide structured editing, and the model uses that to edit.

Please, please turn up the contrast on your darkmode text! The headings are fine, but the paragraph text fails accessibility guidelines badly (and I personally find it very hard to read)

Filed an issue for codex

https://github.com/openai/codex/issues/11601

Great article and tbh I thought it would’ve been implemented that way makes sense to hash and save mainly context I don’t expect them to care about token usage

How about Kimi tho how can I play with it?

This is awesome. Adopted it in my personal pi-mono repo :)

It had some bugs around whitespace replacements but the model seems happy with it now.

Thanks and keep it up! Shoutout to @jahala for tilth as well.

I switched from a basic prompt wrapper to structured tool use with Claude Code and the quality of output jumped overnight. Same model, completely different results.

This is very nicely done. We have seen the same issue at a higher level of getting separators right when generating multiple files in a single inference call.

Arguably I would think that the last year was mainly inner harness improvement instead model improvement but I could be wrong, just feels like that to me

Still weird to me that most people are not just giving an LLM an access to an editor, forcing it to write shell scripts to edit files. Shrug.

Google banning you for benchmarking is crazy, are you sure that's the cause? How would they even know you are benchmarking?

My experience exactly! I’ve recently become so tired of the Claude harness that I switched to OpenCode (which is extremely good compared to Claude). However, OpenCode is also tedious to change, and it inherits all the “good stuff,” like treating agents as Markdown files and all the dancing around with hooks/plugins/skills scattered all over the place. Getting stuck again and again, I’ve ultimately come to the conclusion that this must be solved by writing my own damn coding agent, with extensibility that’s acceptable for real-world engineering.

Great article, recommend reading all of it.

> Why bother, you ask? Opus may be a great model, but Claude Code to this day leaks raw JSONL from sub-agent outputs, wasting hundreds of thousands of tokens. I get to say, “fuck it, subagents output structured data now”.

This is why I find the banning of using Claude subscriptions in other harnesses is so heinous. Their harness that they're forcing onto everyone has tons of big issues including wasting massive numbers of tokens. Very much in line with intentionally refusing to adhere to standards in the most IE6 way possible.

if you want to quickly try it on codex https://github.com/tartavull/hashline

I use small model I like to give them TOC more than lines wonder how it'd stack up with the hashline approach

read_toc tool:

...

  {

    "name": "mcp",

    "qualified_name": "mcp",

    "type": "constant",

    "docstring": null,

    "content_point": "src\\mcps\\code_help\\server.py::17::18::python::mcp",

    "is_nested": false

  },

  {

    "name": "handler",

    "qualified_name": "handler",

    "type": "constant",

    "docstring": null,

    "content_point": "src\\mcps\\code_help\\server.py::18::19::python::handler",

    "is_nested": false

  },

....

update_content tool:

{

  "content": "...",

  "content_point": "src\\mcps\\code_help\\server.py::18::19::python::handler",

  "project_root": ....

}

You should have patented this and you'd be rich.

This sounds like a good optimization task to give a long-running agent. Ask it to come up with experiments and maximize the % of successful edits.

really enjoyed reading this, although I'm a dumb farmer and it took me a while to understand lol

> Why bother, you ask? Opus may be a great model, but Claude Code to this day leaks raw JSONL from sub-agent outputs, wasting hundreds of thousands of tokens. I get to say, “fuck it, subagents output structured data now”.

The VC economics are creating a reality distortion field where Anthropic is incentivized to burn more tokens so they can rent more GPUs so they can get more investment, and where I am incentivized to pipe the LLM inputs into `claude -p` and blast 50KB of useless proompt onto it so they don't ban me from their 95% discounted API endpoint.

Why not just use line numbers?

Is there a skill file I can use for these edits?

Great post!

The logical end state of this line of reasoning is a collective action problem that dooms the frontier lab establishment. You can't devote model capacity to having an attention transformer match nested delimiters or cope with bash and be maximally capable, you can't mix authentication, authorization, control plane, and data plane into an ill specified soup and be secure enough for any that isn't a pilot or toy ever.

If you run this out, you realize that the Worse is Better paradox has inverted, it's an arbitrage, and the race is on.

I agree with this article completely, nice to see it presented quantitatively.

>re "only" the harness changed

In our experience, AI's are like amnesiacs who can barely remember what they did three minutes ago (their last autonomous actions might still be in their context if you're lucky), with no chance at remembering what they did three days ago. As such, the "harness" determines their entire memory and is the single most important determinant of their outcome.

The best harness is a single self-contained, well-commented, obvious, and tiny code file followed by a plain explanation of what it does and what it's supposed to do, the change request, how you want it to do it (you have to say it with so much force and confidence that the AI is afraid of getting yelled at if they do anything else) and a large amount of text devoted to asking the AI not to break what is already working. Followed by a request to write a test that passes. Followed by asking for its judgment about whether it broke what was already working on or not. All in one tiny crisp prompt.

With such a harness, it's able to not break the code one time in twenty. If you use reverse psychology and ask it to do the opposite of what you want, it rises to fifty-fifty odds you'll get what you're trying to do.

Don't believe me? You can watch the livestream (see my previous comments).

Baby steps toward Utopia.