Seed

About

Seed is Vivek Haldar's deliberately minimal agent experiment: one Python file that calls a language model with a single tool, exec, and nothing else, drawing on McCarthy's metacircular evaluator and homoiconicity as its design inspiration. Everything a framework normally provides — tools, memory, skills, conventions — must be grown by the agent itself into a self/ directory it can edit; sessions are otherwise ephemeral, so anything not written to self/ is lost, and verbatim transcripts accumulate in self/sessions/ as a flight recorder the agent may learn to study. Planting via uvx copies seed.py into a fresh directory, germinates self/SELF.md, and initializes a git repo; each directory grows a different agent, making it less a product than a probe into how much structure an agent can bootstrap on its own. Models route through Simon Willison's llm library, defaulting to openai-codex/gpt-5.6-sol with the ChatGPT login from the Codex CLI.