- Published on
Minimal, self-modifying agent harness
- Authors
- Name
- DeepHorus
A minimal, self-modifying agent framework called "Seed" has been released, offering a radical departure from traditional AI agent toolkits. Unlike complex frameworks that provide pre-built tools and structures, Seed starts with a single Python file, seed.py, which forms a small loop connecting a language model to a single bash execution tool. This design philosophy aims to create a true "seed" from which an agent's capabilities can organically grow over time.
The project, authored by Vivek Haldar, is designed to be planted in a local directory using a simple command. Upon initialization, it copies its core files, creates a self/ directory for agent-owned files, and starts a new git history. This setup ensures the agent's core loop is part of its own versioned history from the very beginning. The system prompt, loaded from a file the agent can rewrite, is the only initial guidance provided.
Session interactions are recorded verbatim to self/sessions/*.json files, acting as a "flight recorder" rather than persistent memory. The agent does not automatically access these past transcripts but can develop tools to study its own history if it chooses. All persistent knowledge, skills, and conventions must be actively written and curated by the agent into its self/ directory between sessions.
Seed leverages Simon Willison's "llm" library for model and key management, supporting providers like OpenAI, Anthropic, Gemini, and OpenRouter. The default model is openai-codex/gpt-5.6-sol, but users can specify alternatives like gemini-2.5-pro on a per-session basis. This flexibility allows the core minimal loop to interface with a wide range of powerful language models.
The architecture draws inspiration from conceptual computing principles like McCarthy's metacircular eval and homoiconicity, as detailed in the project's DESIGN.md. The conscious decision to forgo a framework in favor of a minimal starting point is intended to explore new paradigms in agent autonomy and development. Each planted directory creates a distinct agent individual, whose capabilities diverge based on unique experiences.
This approach presents significant risks and challenges, primarily placing the burden of growth and tool creation entirely on the agent itself. It eschews safety guardrails and structured conventions common in other systems. The project positions itself as an experimental platform for studying agent evolution in a sandboxed environment.
Looking ahead, Seed provides a blank canvas for developers and researchers to experiment with truly self-sufficient agents. The future development of any Seed instance is entirely open-ended, contingent on the ingenuity of the underlying model and the directives of the human user. It represents a fundamental rethinking of what constitutes an agent's starting point and path to proficiency.
