EvolvingAgentsLabs/evolving-agents

Your toolkit for autonomous, evolving agent ecosystems. Create, execute, govern, and evolve agents that learn from experience, collaborate, build new capabilities, and operate safely within defined guardrails.

What it solves

This project serves as the original foundation and architectural map for a suite of AI agent systems. It demonstrates the transition from a prompt-based governance model (asking an agent to behave) to a verifiable, constraint-based model (preventing an agent from emitting invalid tokens).

How it works

Originally known as the Evolving Agents Toolkit (EAT), the system consisted of 18,680 lines of Python across twelve subsystems, including a SmartLibrary for versioned components, a SmartAgentBus for routing, and a governance layer called Firmware. It relied on MongoDB Atlas for infrastructure. The project now functions as a reference point to show how these components were rebuilt into more robust, verifiable tools like token-trie (decoder-level constraints) and agentvcs (verified evolution loops).

Who it’s for

Developers and researchers interested in the evolution of AI agent architectures, specifically those moving from naive prompt-based instructions to verifiable software constraints.

Highlights

  • Architectural Evolution: Maps the transition of five subsystems into specialized, verifiable repositories.
  • Governance Shift: Contrasts prompt-based governance with token-level masking to ensure agents cannot produce malformed output.
  • Historical Reference: Maintains the original code to provide a direct comparison between early agentic designs and current evidence-based implementations.