Agent skill · software engineering · aradotso
autoagent-harness-engineering
autoagent-harness-engineering — an agent skill for coding agents.
Why this skill is useful
Adds executable scripts and a complete framework for autonomous agent harness engineering that significantly enhances AI capabilities.
What it needs
Requires docker, python, uv installed locally. Requires openai account access. About 7k tokens when loaded. Last updated 2026-07-09. 70 stars on the source repository.
What this skill does
bash Install uv curl -LsSf https://astral.sh/uv/install.sh sh Clone and install dependencies git clone https://github.com/kevinrgu/autoagent cd autoagent uv sync Configure environment variables cat > .env << 'EOF' OPENAIAPIKEY=$OPENAIAPIKEY Add other provider keys as needed EOF Build the base Docker image docker build -f Dockerfile.base -t autoagent-base . autoagent/ ├── agent.py # Single-file harness under test (primary edit surface) ├── program.md # Meta-agent instructions + directive (human edits this) ├── Dockerfile.base # Base Docker image for task containers ├── tasks/ # Benchmark tasks in Harbor format │ └── my-task/ │ ├── task.toml │ ├── instruction.md │ ├── tests/ │ │ ├── test.sh │ │ └── test.py │ ├── environment/ │ │ └── Dockerfile │ └── files/ ├── jobs/ # Harbor job outputs (auto-generated) ├── results.tsv # Experiment log (created by meta-agent, gitignored) ├── run.log # Latest run output └── .agent/ # Optional workspace artifacts meta-agent reads program.md → inspects current agent.py → runs benchmark → diagnoses failures → modifies agent.py (prompt, tools, config, orchestration) → runs benchmark again → keeps change if score improves, discards if not → repeats bash rm -rf jobs mkdir -p jobs uv run harbor run \ -p tasks/ \ --task-name "<task-name>" \ -l 1 \ -n 1 \ --agent-import-path agent:AutoAgent \ -o jobs \ --job-name latest > run.log 2>&1 bash rm -rf jobs mkdir -p jobs uv run harbor run \ -p tasks/ \ -n 100 \ --agent-import-path agent:AutoAgent \ -o jobs \ --job-name latest > run.log 2>&1 Read program.md and let's kick off a new experiment! tasks/my-task/ ├── task.toml # Config (timeouts, metadata) ├── instruction.md # Prompt sent to the agent ├── tests/ │ ├── test.sh # Entry point — writes /logs/reward.txt │ └── test.py # Verification logic ├── environment/ │ └── Dockerfile # FROM autoagent-base └── files/ # Reference files mounted into container toml [task] name = "my-task" description = "A sample benchmark task" timeout = 300 [task.metadata] cat …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills aradotso/autoagent-harness-engineering