canvas-orgcanvas-orgOfficialยท1 Agent Skills Included

meta-agent

Self-improving agent harnesses through automated experimentation loops

Improves agent performance by automatically rewriting prompts, tools, and control flow based on execution traces. Runs optimization loops that diagnose failures, propose harness changes, and validate candidates against benchmarks. Stores every experiment result on disk so you can audit what worked and what failed. Keeps the base model frozen while learning through versioned code and instructions.
npx skills add canvas-org/meta-agent --all -g -y

All Skills in This Repository (1)

Pure Emerald Level Indicators

Frequently Asked Questions

FAQPage Schema
How to install Meta-Agent?โ–ผ

Run `npx skills add canvas-org/meta-agent --all -g -y` in your terminal to install everything globally. It requires Python 3.11 or later and API keys for OpenAI or AWS Bedrock.

What does Meta-Agent actually do?โ–ผ

It runs your agent on benchmark tasks, records the traces, diagnoses failures, and automatically rewrites the surrounding harness code to improve accuracy. The base language model stays frozen the whole time.

Does Meta-Agent fine-tune the language model?โ–ผ

No. It never touches model weights. All learning happens in the harness code, proposer instructions, and a persistent on-disk experience store.

What benchmarks does Meta-Agent support?โ–ผ

It ships with adapters for tau-bench v3 airline tasks, Plan-RewardBench, and trajectory-level judge evaluations. You can also define your own benchmark with a simple YAML config and adapter.

Can I run long optimization searches in the cloud?โ–ผ

Yes. The repository includes a Modal deployment guide for running extended searches on cloud infrastructure instead of your local machine.

Related Repositories in Education & Research

View All in Education & Researchโ†’