Better Models: Worse Tools

This is easily solved with good error messages.

Claude always gets the syntax wrong on my tool calls.

So I did a revolutionary thing and made the error output print helpful guidance on how to correctly call the tool.

The agent tries again and always gets it right. Total time “wasted”: 1-2 seconds. It happens every session, but it only happens once per context window. After that the agent holds on to the lesson.

To do this for your own tool calls, imagine what you’d do in the agent’s place - what info you’d need so you can correct your mistake. Assume the agent wants to achieve the goal so it’ll try again. These are probabilistic systems, so we need to give them an extra loop to get the deterministic bits right.

When building agent integration for my serverless backend https://saasufy.com/, I decided to not use MCP but to put curl commands inside skill markdown files instead: https://github.com/Saasufy/skills

The curl command is extremely popular so models seem to be really good at using it.

Also I like that curl uses a bash syntax and my platform requires JSON payloads; it makes the separation clear to the agent. I find it to be very reliable.

As critical as I am about articles endlessly concerned with the weaknesses of closed-source cloud LLMs, this one is pretty great, and not just because it concerns interactions with Pi, which looks to me like it's going to end up a sort of quasi-reference implementation of an open source harness, and because it has so much useful technical detail.

But:

"Now I’m somewhat worried about the track we’re on here. Alternative tool schemas might not just be unfamiliar. They might be implicitly punished by post-training that optimizes for one particular, forgiving tool ecology."

Only implicitly?

--

Many decades ago when I was working on research related to using MOOs as a learning environment, you would add "tool calls" into the stream of text that a MOO object might generate, so your rich client would e.g. show a picture, load a web page in a frame, move you on a map, trigger a change in an on-screen representation of an object.

Everyone who tried this in MUD/MUSH/MOO clients ran into more or less the same problems that LLM clients do: any attempt to shoehorn control sequences into in-band content was riddled with security risks, objects accidentally triggering the wrong interface etc.; you could never truly communicate out-of-band.

The more I read about how agentic harnesses work, the less embarrassed I feel about the code twenty-something-year-old me wrote in a MOO client.

> You can ask the model to produce valid JSON

Doesn't always work, for better performance you can kneel and start begging

It's not the failed call that worries me. The call itself was correct, and the only thing off was a couple of invented fields. That makes the runtime feel like part of the model's interface rather than just an implementation detail. Train a model in a forgiving environment and other runtimes end up inheriting its habits.

It sounds like harnesses might have to start to have model by model system prompts, though retrying works, I guess. It reminds me of the ancient times when browsers all read HTML and CSS differently, and differently on different devices. In that sense, this is nothing new. I was going to say, at least we don't have different device types, but then, the model still has to output the right variant of `grep` as well.

Suggestion for Pi: capitalize tool names for the Sonnet/Opus models (edit -> Edit, bash -> Bash, ...).

The rationale: Anthropic's own harness (Claude Code) uses PascalCase tool names — Bash, Edit, Read, Write, Glob, Grep. Since the models are post-trained/aligned against that harness, those naming conventions are effectively baked into the model. Matching your harness's tool names to the same casing puts your inputs closer to the training distribution, which lines up with the more reliable tool use I've seen in evaling.

A related pattern that fits the same distribution: for long outputs, have the model reserve placeholders first and complete the work across multiple steps.

Reference: https://github.com/evotai/evot/commit/765151796c43965964a9da...

There's a spectrum of possible explanation, from "this is a model training artifact which for now they correct via the harness" through to "this is deliberate, and creates a constantly moving target to make third-party harnesses less efficient for lock-in purposes".

I'd not discount the adversarial end of the spectrum.

Once models get better, we could avoid paying for a cache read on edit or write calls, and have the model assume they succeeded and not interrupt the stream to get output. We can then just parse the output and once we encounter such a silent toolcall execute it. With high probability its correct (glm in pi for me had 95% tool call success rate) and we can continue, else rewind. As a workaround, you dont want to use the provider feature that interrupts the stream after a tool call, but instead parse the reasoning. I tried this in pi and it kind of worked, but the model got confused about whether edits had been applied and in several runs either double checked or used the bash tool instead, negating any possible benefits.

Surprised models still output tools as text when for ages we’ve been able to constrain the output at the inference engine level and constrain the model what tools, parameters etc are available

Edit: found it, it’s called Grammar-Constrained Decoding (GCD)

Pi is my daily driver. I noticed the same phenomenon, and had Claude analyze all my past transcripts for classes of 'edit' error. Built an extension which patches the edit tool to self-heal on the majority of those kinds of calls. It's not 100%, but it cuts down on the rejections quite a bit and saves a few round trips.

EDIT: It's still quite fascinating seeing the kinds of things the models keep trying to do. It almost seems like when a human has slightly off with their nervous system. The conscious brain wants to do one thing, but for some reason the signals aren't getting to the hands correctly.

> In case you are curious about Fable: I intentionally did not test it because I was not sure if the classifiers they are running might downgrade me to Opus silently.

Is this still a thing? I thought Anthropic walked back the silent downgrades so now all the different domains downgrade non-silently.

A better solution might be not to constrain the generation, but to remove invalid fields from the tool call in the assistant message. So on the next turn, the model receives chat history which contains it's tool call, but without extra arguments. You can do that in OpenAI chat completions / responses, not sure about Anthopic API.

There is still a downside, sometimes the model really wants to include an additional argument for whatever reason it reasoned towards, and it needs the error message to understand that the argument doesn't exist. Otherwise if the argument is manually removed from it's tool call, the model will think that it accidentally left out the argument and start retrying and might go into a loop.

In my harness i implemented apply_patch just taking unified diffs for patch -p1. I was shocked to see how bad models are at generating them. I started logging diff failures to analyse -

- All models are terrible at generating line numbers for a proper diff, give up on them

- Some models (Owl-alpha) must have been post-trained on Codex transcripts, because they occasionally push its V4A patch format into any diff tool available

- Codex puts a lot of info in its system prompt about the desired patch style, making larger hunks instead of granular ones, etc

For the smaller local model I use (DS4 Flash) I just modified the tools to match the common error case of the model. If the mountain won’t come to Mohammed…

LLM can write programs in any programming language it knows about. So how about askinng it to write a shell-program that does the tool-calls on the client?

You might want to run in some kind of sandbox to prevent the LLM from taking over the world, security is an isssue. But apart from that why not make the LLM write shell-programs instead of relying on JSON etc. ? Shell-scripting is the language for controlling the OS.

It's been clear for some time that model tool calling is heavily fit to a few common patterns, it's unsurprising that a tool call that looks the same or has the same name, but works differently, is falling back to priors and causing problems.

Things are not quite AGI yet; which is why people are now saying that intelligence is the harness + model, because the harness makes up for limitations in generalization.

My favorite feature from Claude code is the "auto" mode to dynamically approve permission queries that are reasonably safe. Unfortunately the standard pi sandbox extension doesn't support it. Pi should really build permissions into the agent leve.

https://github.com/carderne/pi-sandbox

Different but related: When you use a Codex subscription in an agent like Pi or OpenCode, all the requests and tool call execution go through a sandbox owned by Codex app server, and all the tool calls function somewhat differently, and you can't read files outside of the sandbox as easily. It's currently tripping me up a bit when building a model router.

This has been the case since the early days. Aider had a bunch of code to be very forgiving with formatting of tool calls (file editing in particular at first). It's just the nature of the beast. It surprises me that Pi doesn't have a lot of this kind of stuff built in too

Yep. I spotted the same thing in piclaw (which relies on the pi runtime) but did not have time/energy to do a lot about it—and fable does the same, as far as I can tell, with one out of five or six edits failing. But I prefer OpenAI models for coding, so it wasn’t a real problem.

This makes sense to me, much as I don't like it. IMHO the strategy taken by StrongDM's attractor coding agent seems like a path of least resistance. Directly target the LLM providers APIs and directly target their default tools.

Does Pi even need read/write/edit tools? Couldn't it just have bash commands and get the model to use e.g. sed for everything?

I suspect this isn't a malfunction, but rather a deliberate measure designed to counter so-called distillation attacks.

I want dark window chrome and light contents but browsers seem completely unwilling to let me have this option.

building deterministic tools on non-determinism is hard enough; try adding another layer where your cloud provider decides to massage the context, realigns it's permitted output, arbitrarily downgrades context to cheaper models, or they hire an MBA who determines your plan value can be tied to a degraded model under a new shrinkfied.

It's amazing anyone watched the last 2 decades of tech's enshitification and wants to hook their wagon to this shitshow.

> [...] newer Claude models sometimes call Pi’s edit tool with extra, invented fields in the nested edits[] array

> My strongest hypothesis is that this is not random deterioration but a training artifact. [...] Anthropic’s own client appears to expect and accept a fair amount of slop and repairs it, mostly silently

> If reinforcement learning happens in a harness like that, or a simulation of one, then slightly malformed tool calls can still complete the task and receive reward.

> Worse, the model may become very strongly adapted to the canonical Claude Code edit tool shape.

> Tool schemas are somewhere in the distribution and some shapes are close to what the model saw during post-training and some are far away.

Great article.

Interesting root cause hypothesis. Couldn't one simply strip the slop-handling from the RL env's harness to avoid this though?

I do agree on the walled garden being built here. Proprietary frontier models performing best in proprietary harnesses makes sense for Anthropic's interests.

I guess we are going to get even more of this. Where models and tools start producing nonsensical results and no-one understands why it appends and we must read articles like this that catch it.

closed source harness + RL fine tuning on customer prompts on said harness is becoming a kind of economic moat (?)

We’re entering the era of AI trained by previous generations’ slop. It’s not surprising that it’s sloppy.

Open source developer surprised and concerned by the trajectory their favorite proprietary software is taking.

Hey, an article right up my alley! AI infrastructure/tools engineer here (hic-ai.com); my flagship product, HIC Mouse, is a precision-editing system for coding agents designed to work across a wide array of models and harnesses. Mouse provides 11 tools exposed via MCP for read-, find-, and edit-operations, using a coordinate-based schema (as well as exact and multiple string replacement), a Dialog Box inspect/refine/save/cancel changes functionality controlled by the agent to force staging and review of multi-operation or large edits before changes are written to disk, and extensive agent guidance mechanisms or guardrails to help the agent realize if it's about to do something potentially destructive or overly verbose.

I definitely think models may be trained to use particular popular harnesses or expect certain fields in the editing-tool or other tool schemas. Rather than trying to conform to (or force) one particular format, my approach instead is to design flexibly enough to handle a wide array of possible inputs and tool calls, but that also help the agent recover whenever its tool calls truly can't be salvaged and have to return etrors, and to auto-normalize results whenever reasonable to do so. It really does make a very dramatic difference (I wouldn't have bothered to launch if I thought it wasn't a meaningful advance) but anyway, just wanted to share my perspective given that I live and breathe this problem all day, every day.