Vivek Haldar Releases Seed: A Minimalist Self-Modifying Agent Harness
Rejecting framework bloat, Seed provides a bare-bones loop that allows AI agents to autonomously build their own tools and memory.
Vivek Haldar has released Seed, a minimal agent harness designed as a foundational starting point for AI agents to grow their own capabilities. The project represents a shift away from pre-configured AI frameworks toward a model of autonomous evolution.
At its core, Seed consists of a single file, seed.py, which establishes a basic loop connecting a Large Language Model (LLM) to a single bash 'exec' tool. To manage models and API keys, the harness utilizes Simon Willison's 'llm' library, providing compatibility with providers including OpenAI, Anthropic, Gemini, and OpenRouter. Unlike traditional agents with hard-coded instructions, Seed stores its system prompt in a file that the agent can rewrite, allowing it to modify its own behavior over time.
All persistent data, tools, and skills must be developed by the agent within a dedicated 'self/' directory; otherwise, sessions remain ephemeral. For auditing purposes, every session is recorded as a JSON transcript in 'self/sessions/', though the agent does not load these transcripts at boot by default.
A Counter-Trend to Framework Bloat
Seed is positioned as a direct response to the increasing complexity and "bloat" found in modern AI agent frameworks. Rather than providing a suite of pre-built plugins and complex orchestrations, Seed draws inspiration from homoiconicity and McCarthy's metacircular evaluator. This approach treats the agent as an evolving entity rather than a static tool, stripping away the engineering layers that typically dictate how an agent should function.
Shifting the Development Paradigm
This minimalist architecture shifts the paradigm of agent development from engineering a rigid framework to "planting a seed." By providing only the absolute minimum viable harness, the system allows researchers and developers to observe how an agent autonomously organizes its own memory and toolset. This organic growth potentially leads to agent architectures that are less biased by human design and more adapted to the actual tasks they perform.
The Path to Autonomous Growth
As the agent operates, it is expected to write its own scripts and memory structures into its directory, effectively building its own operating system. The primary point of interest for observers will be how these agents prioritize which tools to build first and how they manage the stability of their own system prompts. While the core logic remains simple, the resulting emergent behaviors remain the key metric for the project's success.