PenguinHarness

PenguinHarness is an open-source framework that lets you build AI agents using an AI agent. Instead of writing chains, tools, and prompts line by line, you give the harness a single sentence and it generates a complete, runnable agent application — scaffold, code, and instructions included. The tool runs locally on your machine with a bundled Node runtime, keeping all data in a local directory. It supports over 1,000 models and can run on as little as a single CPU.

What makes PenguinHarness stand out is its recursive self-improvement loop. The system uses multiple evaluator agents to score a target agent, then an optimizer agent analyzes the scores and traces to find where points were lost and upgrades the agent from version N to N+1. Every round is snapshotted, so you can roll back if a change regresses. The framework also includes full tracing of every model request and tool call, approval gates before tool execution, credential isolation, and progressive loading that reads content on demand instead of dumping everything into context. Benchmark results show comparable accuracy to Claude Code and OpenAI Codex at a fraction of the cost — as low as 1/70th.

PenguinHarness is built for developers and enterprises who want to automate agent creation and maintenance. Use cases include building RAG apps from documentation URLs, creating coding assistants that improve over time, and running complex data analysis tasks. The one-line install works on Linux and macOS, and the web interface runs locally on port 7364. It's a practical tool for teams that want to move from hand-crafting agents to letting agents build agents.