Every agent framework I’ve evaluated is a race toward more: more built-in tools, more memory abstractions, more orchestration scaffolding you didn’t write and can’t fully see. Seed runs the opposite experiment — it ships with almost nothing and dares the agent to grow the rest.

The frozen core is one file, seed.py: a loop wiring a model to a single tool (bash exec), with its system prompt loaded from a file the agent is allowed to rewrite. Everything else — tools, skills, memory, conventions — has to be grown, session by session, into a self/ directory that is the only thing surviving between runs.

• 🎯 One tool, not fifty — bash exec is the entire action space; the agent builds up from there or it doesn’t build at all
• 🧬 The prompt is mutable state — the agent owns and rewrites its own system prompt, which is either the whole point or the whole risk
• 📼 Sessions as a flight recorder — every run is dumped to self/sessions/*.json, available for introspection but never auto-loaded
• 🌱 Git as the memory substrate — each agent is its own repo, so its evolution is a commit history you can actually diff
• ⚠️ The caveat is the thesis — the authors admit it trusts the agent with capability expansion; that’s the experiment, not a bug

In production I do the exact reverse of this: fixed tool registries, hard guardrails, and observability precisely because I don’t want the agent rewriting its own contract. But the flight-recorder-plus-git pattern is quietly the most interesting part — it’s a cleaner audit trail than most enterprise agent platforms ship. The HN discussion splits between “elegant” and “loaded footgun,” which is usually where the good ideas live.

My bet: almost nobody runs this in production, and almost everybody borrows the git-tracked self/ directory as an agent-memory pattern within a year.

tags: [ agentic-ai ] [ llm-ops ]