Seed: A Minimal, Self-Modifying Agent Harness Released

Introduction to Seed

Seed is a newly introduced minimal agent harness. It functions as a small loop connecting a language model to a single tool, 'exec', which runs bash commands. The system prompt is loaded from a file owned and rewritable by the agent itself.

Self-Modifying Capabilities

Unlike traditional frameworks, Seed requires agents to grow their own tools, memory, skills, and conventions into a dedicated 'self/' directory during each session. This design allows for self-modification and adaptation, as the agent's core loop is part of its individual history.

Each instance of an agent planted in a directory is unique and diverges based on its experiences, fostering individual development rather than relying on a shared framework.

Operational Details

To start, users create a directory and initialize Seed, which copies 'seed.py' and 'run_seed.sh' into it, germinates 'self/SELF.md', and commits these files to a new git repository. Subsequent sessions can be resumed using 'run_seed.sh'.

A verbatim transcript of every session is recorded to 'self/sessions/*.json', serving as a flight recorder. While not loaded at boot, agents can develop tools to analyze their past interactions.

Model Integration

Seed handles models and API keys through Simon Willison's 'llm' library. The default model is 'openai-codex/gpt-5.6-sol', which uses the ChatGPT login from the Codex CLI. Users can override the model for specific sessions.

Bundled providers include OpenAI (via Codex subscription or API key), Anthropic, Gemini, and OpenRouter, which offers access to hundreds of models with a single key.