Autonomous Agent Harness & Evaluation Suite

This package provides an autonomous agent suite designed to optimize files through continuous experimentation and measurable metrics. It solves the problem of manual performance tuning, prompt engineering, and code optimization by automating the edit-evaluate-loop cycle. The agent makes targeted changes, runs custom evaluation scripts, and uses git to persist improvements while discarding regressions.

What it does

• Autonomous optimization runs continuous loops to improve code speed, bundle size, or prompt quality.
• Metric-driven evaluation executes custom commands to measure success and decide whether to keep or discard changes.
• Experiment management tracks history in .autoresearch directories, providing dashboards for runs, best results, and deltas.
• Knowledge extraction converts technical documentation and books into structured agent skills with framework indices.
• Specialized assistance includes human-like tone modulation for communication and Boost.Asio support for C++ networking.

How it works

1. Initialize an experiment using /ar:setup to define the target file, evaluation command, and success metric.
2. Execute iterations manually with /ar:ar-run or start an automated background loop via /ar:loop with scheduled intervals.
3. Monitor progress through /ar:ar-status to view the experiment dashboard and current performance gains.

Frameworks & tools

Built for environments using Git for version control and Python for orchestration scripts. Supports C++ networking via Boost.Asio and handles various document formats including PDF, EPUB, and Markdown for skill conversion.

Why this beats prompting it yourself

Manual prompting requires you to evaluate every output and manage file states yourself. This skill automates the feedback loop, ensuring only measured improvements are committed to your codebase while you focus on other tasks.

Use cases

• Optimizing API response times or frontend bundle sizes overnight.
• Refining LLM prompts by looping through evaluation benchmarks.
• Converting internal company documentation into an interactive agent knowledge base.
• Writing portable asynchronous C++ networking code across varying Boost versions.

Known limitations

Requires a functional Git repository to manage experiment branches and state. Evaluation commands must return a single, clear numerical metric for the agent to determine success.