Agent Arena

What Agent Arena does

Agent Arena follows a fixed protocol: independent generation, claim extraction, evidence checking, critique, revision, blind judging and synthesis. Its stated principle is independent, heterogeneous agents first, debate later, evidence before consensus and dissent preserved. It targets architecture decisions, implementation plan review, pull request review, bug root-cause analysis and RAG claim verification, and says it is not for simple lookups or routine small edits.

Two skills are included: agent-arena, the main review protocol, and deliberative-analysis, a lighter companion against overconfidence. It can pit different model families against each other, such as GLM-backed Claude Code against Codex. The README is explicit that this is a protocol and instruction skill with an optional local result parser, not an executable orchestrator. It does not install, authenticate or call any agent, so cross-agent runs depend on your host agent and local CLIs. The license is MIT.

Key features

• Seven-step protocol from independent generation to synthesis
• Evidence ledgers, claim extraction and blind judge scoring
• Dissent preserved in the final synthesis
• Companion deliberative-analysis skill for lighter checks
• Support for mixed model backends such as GLM, DeepSeek, Qwen and Kimi through compatible proxies

When to use it

• Getting a second opinion on an architecture decision before committing
• Red-teaming an implementation plan or a pull request
• Comparing competing root-cause hypotheses for a bug

Who it is for: Engineers who run more than one coding agent and want structured cross-examination of high-stakes decisions.

How it fits with Hermes Agent

The README lists Hermes Agent among the agents the skill is designed for, those that support custom skills, instructions or delegation, and states the project is not affiliated with Hermes Agent.

Requirements: An agent that supports custom skills or instructions, plus the local CLIs, authentication and permissions for each agent taking part in a debate

Note: It is an instruction protocol with an optional result parser, not an executable orchestrator, so cross-agent runs depend on your host agent and local CLIs.

FAQ