canvas-org

canvas-org / meta-agent

Public

Continual harness optimization

agent anthropic claude harness llm
28
1
100% credibility
Found Apr 07, 2026 at 28 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

A framework that automatically optimizes AI agent configurations by analyzing execution traces to improve performance on benchmarks like customer service simulations.

How It Works

1
📰 Discover the booster

You hear about a handy tool that automatically makes your AI assistant smarter at solving real-world problems like customer service tasks.

2
📝 Set up your challenges

You describe the tasks you want your AI to master, like fixing billing issues or handling airline bookings, with simple checks for success.

3
🔗 Connect your AI helper

You link the smart AI service that thinks and acts, so everything is ready to go.

4
📊 Test the starting point

You run a quick check to see how well your AI does right now on your tasks.

5
🚀 Launch the improvement magic

You start the special learning loop that watches what goes wrong and creates better plans automatically.

6
📈 Watch it get better

Over a few rounds, it tries new ideas, learns from each try, and boosts success rates step by step.

7
🏆 Pick the champion setup

You review the results, compare the winners, and choose the best version that solves the most problems.

🎉 AI agent supercharged

Your assistant now handles tough tasks much better, turning 67% success into 87% or more effortlessly.

Sign up to see the full architecture

6 more

Sign Up Free

Star Growth

See how this repo grew from 28 to 28 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is meta-agent?

Meta-agent automates optimization of AI agent harnesses—system prompts, hooks, and options for tools like Claude Code—using execution traces from failed runs. Developers define tasks in YAML with instructions, workspaces, and verify commands, then run a baseline eval and an outer loop that proposes and tests improvements, jumping tau-bench scores from 67% to 87% without gold labels. Built in Python, it hooks into Anthropic and OpenAI APIs via the Claude Agent SDK, fitting into continual AI GitHub trends like continual learning surveys and meta-agentic AI experiments.

Why is it gaining traction?

It delivers label-free gains on real agent benchmarks like tau-bench (airline/retail customer service), with CLI tools to list candidates, diff results, and inspect failures—making iteration transparent and fast. Unlike manual tuning, the proposer reads traces and evolves configs in a continual RL-like loop, supporting holdouts to avoid overfitting. Early users get reproducible boosts on custom tasks, echoing continual pretraining GitHub repos.

Who should use this?

Agent engineers tuning Claude setups for coding or customer service sims, especially on tau-bench or YAML-defined evals with verify scripts. Teams in meta-agents research environments or continual world GitHub projects experimenting with harnesses for games like Valorant agents. Skip if you're not using Anthropic models or lack API keys.

Verdict

Try it for agent eval loops—solid quickstart and CLI make it practical despite 28 stars and 1.0% credibility signaling early maturity. Docs are benchmark-focused but light on edge cases; pair with your own tests for production.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.