Agent Harness Lab
An open-source experimental tool for providing a reproducible evaluation and comparative testing process for improvements to coding-agent instruction files (such as AGENTS.md) and skills. It supports Claude Code and Codex.
Timeline (last 90 days)
The open-source tool Agent Harness Lab aims to provide a repeatable testing process for instruction optimization of coding agents.
Agent Harness Lab is currently experimental, supports Claude Code and Codex, and is open-sourced on GitHub (https://github.com/mihazs/agent-harness-lab).
Agent Harness Lab's workflow includes: inspecting local execution evidence and evaluation failures, automatically identifying improvement points, staging project instruction and skill changes, comparing candidate solutions with baseline behavior, and reporting results and limitations.
The key to Agent Harness Lab's workflow is that it does not require pre-defining failure descriptions, lowering the barrier to use.
Agent Harness Lab emphasizes that behavioral changes are not adopted automatically and require additional certification safeguards.
Users can run the command `ahl analyze /path/to/project` for analysis; this command only performs analysis and makes no changes to project files.
The maintainer of Agent Harness Lab explicitly states that the project is currently experimental and does not claim to have proven improvement effects on coding tasks.
In a Reddit post, the maintainer asked developers: how do you currently decide whether a new instruction or skill is worth keeping?
All Facts (5)
The open-source tool Agent Harness Lab aims to provide a repeatable testing process for instruction optimization of coding agents.
50%UnverifiedAgent Harness Lab's workflow includes: inspecting local execution evidence and evaluation failures, automatically identifying improvement points, staging project instruction and skill changes, comparing candidate solutions with baseline behavior, and reporting results and limitations.
50%UnverifiedThe key to Agent Harness Lab's workflow is that it does not require pre-defining failure descriptions, lowering the barrier to use.
50%UnverifiedAgent Harness Lab emphasizes that behavioral changes are not adopted automatically and require additional certification safeguards.
50%UnverifiedUsers can run the command `ahl analyze /path/to/project` for analysis; this command only performs analysis and makes no changes to project files.
50%