EvolvingAgentsLabs/evolving-agents

What it solves

This project provides a versioning and evolution system for AI agents, specifically as a plugin for the Claude Agent SDK. It addresses the gap where the SDK allows forking sessions but lacks the ability to merge, diff, or verify that an agent's evolution (changes to its code, goals, or models) has not regressed in performance.

How it works

The project implements a version control system for agents (agentvcs) that treats an agent's configuration—including its code, goals, models, and traces—as a single unit (a commit). It provides tools to:

• Snapshot and Version: Use avcs_commit to save the state of an agent.
• Compare: Use avcs_diff to compare two snapshots across different dimensions.
• Merge: Use avcs_merge to rejoin two snapshots, including a reconciliation process for goals and traces.
• Verify: Use avcs_freeze to prevent the deployment of an agent if its evaluation fails.
• Secure: Use Ed25519-signed commits to ensure transcripts have not been edited.
• Preserve: Use a PreCompact hook to archive traces before they are summarized away by the SDK.

Who it’s for

Developers building complex AI agents that require rigorous versioning, regression testing, and the ability to merge changes from different experimental branches of an agent's evolution.

Highlights

• Claude Agent SDK Integration: Functions as a local plugin with zero runtime dependencies (standard library only).
• Agent Version Control: Introduces the concept of "commits" for agent state, enabling diffing and merging of agent configurations.
• Regression Prevention: The avcs_freeze tool ensures that agents cannot be "shipped" if they fail evaluations.
• Evidence-Based Development: The repository includes a detailed "Evidence" section documenting failed architectural experiments with dual-embedding resolvers and memory hierarchies, promoting a transparent development process.