evolving-agents
A director agent that learns how your team works: it plans by walking your CLAUDE.md and memory as linked statements, delegates to Claude Code or local models,
What is evolving-agents?
A director agent that learns how your team works: it plans by walking your CLAUDE.md and memory as linked statements, delegates to Claude Code or local models, and keeps each lesson as a gated, versioned patch. Measured: 0.929 vs 0.589.
Overview of evolving-agents
What it is
The director is an agent that learns how a team works. It plans by walking the team's CLAUDE.md and memory as linked statements, delegates tasks to Claude Code or local models, and keeps each lesson as a gated, versioned patch. It ships as a Claude Code plugin and is built to move to Anthropic's Managed Agents and Google's Antigravity agent. It provides an auditable external memory of how you work — methods and operational knowledge, learned only through a gate, kept as plain files.
Key features
- Recalls the rule at the moment it applies and cites the statement in every step of its plan.
- Learns from corrections: each lesson is proposed, checked against existing knowledge, and kept only if accepted as a reversible, versioned patch.
- Learns where work should be done: moving a task from the frontier to a local model once causes the next similar task to be planned there.
- Rules that must hold are enforced as code, not requests (e.g., confidential tasks never leave for a lane not cleared).
- Auditable memory: every lesson is a diff of When/Do/Because statements with probes, lint, optional change check, and a recorded decision.
- Portable: the memory is Markdown; other providers' models follow it as plain notes (Gemini measured at 0.875).
- Cheaper: with Opus planning, 34% less cost per plan for the same score; a local model can plan from recipes at no cost.
Use cases
- Teams that maintain a CLAUDE.md, AGENTS.md, or Claude Code's memory and want rules to be followed consistently under pressure.
- Ensuring that corrections and lessons are captured, gated, and reversible.
- Delegating tasks to appropriate models (frontier or local) based on learned preferences.
- Maintaining an auditable, portable external memory of working methods that can be used by other models.
Getting started
As a Claude Code plugin: /plugin install director --marketplace EvolvingAgentsLabs/evolving-agents /director:compile # your CLAUDE.md, AGENTS.md and memory → linked statements /director:reflect # after a task: correction becomes proposed edit, kept only on your yes /director:recipes # optional: recipes a local model follows to plan from memory Or the harness with two-lane interface: bash git clone https://github.com/EvolvingAgentsLabs/evolving-agents && cd evolving-agents/packages/director python3 -m director.compile --claude-md ~/your-project/CLAUDE.md --memory-dir ~/.claude/projects//memory --workdir ~/your-project python3 -m director.server --config ~/.config/director/director.json # open http://127.0.0.1:8740 Needs Python 3.10+ and Claude Code (claude on your PATH). Standard library only.
License and pricing
License: Apache-2.0. Open source.
Based on the project's public README and repository metadata.