AlphaEval: Evaluating AI Agents in Production

Existing agent protocols (like A2A and MCP) leave lifecycle management, version tracking, and safe evolution underspecified — leading to brittle, monolithic systems.

Autogenesis Protocol (AGP) fixes this by cleanly separating what evolves from how it evolves:

1.⁠ ⁠RSPL (Resource Substrate Protocol Layer): Treats prompts, agents, tools, environments, and memory as versioned, protocol-registered resources with explicit lifecycle and state management.

2.⁠ ⁠SEPL (Self-Evolution Protocol Layer): Provides a closed-loop operator interface (Reflect → Select → Improve → Evaluate → Commit) with full auditability, lineage tracking, and rollback safety.

Built on AGP, the Autogenesis System (AGS) is a multi-agent framework that dynamically instantiates, retrieves, and refines these resources during execution — turning static agents into continuously self-improving systems.

Results (across challenging benchmarks requiring long-horizon planning and tool use):

1.⁠ ⁠GAIA: 89.04% accuracy (beats strong baselines)

2.⁠ ⁠GPQA-Diamond: up to +21% over gpt-4o

3.⁠ ⁠AIME math: up to +71% for weaker models

4.⁠ ⁠LeetCode: +10–27% pass rate + significant runtime gains

Consistent, compounding improvements across models and tasks — showing the power of protocol-level self-evolution.