EvolvingAgentsLabs/evolving-agents

What it solves

This project addresses the gap in the Claude Agent SDK where forked sessions cannot be merged, diffed, or versioned. While the SDK allows branching a session, it lacks a mechanism to track changes in an agent's configuration (code, goals, and models) alongside its conversation history, and it provides no way to prevent the deployment of agents that have regressed in performance.

How it works

The project implements a version control system for agents (agentvcs) that treats agent configurations and their traces as single units called commits. It provides a set of tools to manage these versions:

• avcs_commit: Snapshots the agent's code, goal, and models together with the trace.
• avcs_diff: Compares two snapshots across different dimensions.
• avcs_merge: Rejoins two snapshots, including a reconciliation process for goals and traces.
• avcs_freeze: Prevents an agent from being shipped if its evaluation fails.
• Security: Commits are signed using Ed25519 to ensure transcripts cannot be edited.

It is delivered as a plugin for the Claude Agent SDK via an MCP server and three hooks.

Who it’s for

Developers using the Claude Agent SDK who need rigorous versioning, auditing, and evaluation-based deployment gates for their AI agents.

Highlights

• Agent Version Control: Introduces a system to commit, diff, and merge agent states.
• Evaluation Gates: The avcs_freeze tool ensures that only agents passing specific evals are promoted.
• Zero-Dependency MCP Server: The plugin's MCP server is built using only the Python standard library.
• Experimental Evidence: The repository includes detailed measurements of memory architectures and retrieval strategies, documenting both successes and failures to avoid redundant architectural mistakes.