A seed agent is the smallest starting point from which an agent can grow. The entire frozen layer is seed.py—a small loop that connects a language model to exactly one tool (exec, which runs bash) and loads its system prompt from a file the agent owns and may rewrite.

Everything an agent normally gets from a framework—tools, memory, skills, conventions—must instead be grown by the agent, session by session, into its self/ directory. Plant uvx: `mkdir my-agent && cd my-agent; uvx --from git+https://github.com/vivekhaldar/seed.git seed`. First run copies seed.py and run_seed.sh, germinates self/SELF.md, and commits them to a fresh git repo, then drops you into a REPL.

Everything the agent wants to keep is written to self/. Sessions are fleeting; a transcript is recorded to self/sessions/*.json after each turn, acting as a flight recorder. The agent never loads this at boot but can study its past or grow tools to read it. Each directory you plant yields a different agent that diverges with experience.

Configuration is handled by the LLM (Simon's library). The default model is OpenAI's openai-codex/gpt-5.6-sol; you can override it with `./run_seed.sh -m gemini-2.5-pro`. Bundled providers include OpenAI, Anthropic, Gemini, and Open Router, with one key unlocking many models. The design draws on McCarthy’s metacircular eval, homoiconicity, prior art, and accepted risks, detailed in docs/DESIGN.md.