Loading video...
Video Failed to Load
Introducing Unharnessed, a new generation of harness. One tool: a Python REPL. Designed for the next generation of models. On Terminal-Bench 4.0, Unharnessed achieves a new state-of-the-art result.
115,311 views • 1 day ago •via X (Twitter)
44 Comments

With Opus 5.5 on max, Claude Code gets 64.85% on Terminal-Bench 4.0 and costs $14.30 per task. Unharnessed gets 71.2% and costs $12.19. The model is the same, only the harness changes. You can read every turn of the run here:

If you watch the latest models work, they use the predefined tools less and less. They do most of the job in Python, through Bash. If that's what the model wants to do, we might as well give it the best version of it: one tool, a persistent Python REPL, where its variables and functions stay alive between calls. So instead of one call per action, the model writes a script that reads, searches, edits and checks in a single call. For comparison, Copilot hands it 48 tools, Claude Code 46, Codex 31.

Unharnessed isn't just for code. This paradigm beats the classic harness across the board: 3D, maths, data analysis, science and more. Here it is on a 3D model in Three.js, same prompt and same model as Claude Code: 94 turns, 65 minutes and $14.45, against 102 turns, 72 minutes and $21.67.

It's free and open source. You sign in with the subscription you already pay for (Claude, Codex, GitHub Copilot, Google AI or Grok), so there's no API key to set up. Everything here :

*only mac for now, Windows support and I Linux soon! Next week

Every few weeks, someone rediscovers prime-agent from @PrimeIntellect

@PrimeIntellect Its amazing !

seems familiar

Yep but it doesnt save the python variables for the agent

Population of accidentally wiped drives increases.

I am guessing this has no guardrails ?

Don’t claude ToS forbid you from using other harnesses out of CC under sub? We should start comparing subs costs vs api costs Using whatever lighter harnesses that would force API use (ToS) means anyway 10x and more cost for claude usage vs same on sub

Bravo, mais je vois que c’est un fork de Prime Agent, qu’apporte t-il de plus ?

Je vois, interessant !

Un seul outil REPL à la place d'une pile de tools, le contexte doit respirer. Ça tient sur des tâches longues ?

ahead of the game, models will have functionality baked in.

The video looks really good. How did you make it?

Hey, fable 5.2 without any framework

Can we see the prompt? 👀

Just made a video about X topic haha withb3d Motion nothing more.

Crazy.

@delusionalsilex La ptite vidéo bien propre

python repl only. less to break, i like that

Interesting, I was looking into this, I think there are much better implementations out there that involve repls or 1 tool that calls others, just has to be built.

Quel est la différence avec prime agent ?

Nice je vais tester ça. Je te donnerais des nouvelles. Je l’utilise tous les jours. Je me doutais que tu voulais aborder la même philosophie. Je te ferais un retour par MP

What is your sound design pack? 🤤

a python repl as the only tool is bold. kinda like it

I will try it. The promise is interesting, congrats to get your product out there! I am interested also by what @theo would think about it?

Great work Rayane, i'll test that 👌🏼

Sinew 2.0

U is giving ubiquity

Stripping the harness down to one tool is a bold design choice.

one tool is the right bet. the catch is that all that value lives as variables inside one process, so a crash at turn 80 makes the model rebuild state it already had. a persistent repl is only persistent while the process is.

this is so interesting, @theo your take on this?

The recovery case is interesting here: a script edits three files, then fails on the fourth action. What does the next turn know has already happened? Persistent variables help; an equally clear record of completed side effects would make the one-tool design much easier to trust.

Super interesting, will try!

@ZakShark Insane job, will definitely give it a try

Looks interesting

I love the idea of letting the model building it’s own tools. But what if you need to redo a task ? Should you expect it to find back the same trajectory it did before from ex nihilo ?

So it's Prime Agent?

interesting idea! btw cleanest site i have seen in a while

I want this for powershell

always heard that claude models perform better on other harnesses than their own , will check out on this !



