<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> onevcat/argue is a harness-agnostic orchestration package for multi-agent consensus workflows. It addresses the limitations of single AI agents by enabling multiple agents to analyze a problem, debate claims, and converge on a consensus through voting, yielding higher-quality, more rigorous results. It benefits users by reducing hallucinations and providing a structured, verifiable output.

How It Works

The engine orchestrates a debate among AI agents through distinct phases: Initial claim proposal, multi-round Debate with peer judgment and claim merging, and a Final Vote. Agents are scored on correctness, completeness, actionability, and consistency via peer review. The highest-scoring agent acts as the "representative" to compose a human-readable report and can optionally execute a post-debate action. This iterative, cross-examined approach ensures claims survive scrutiny, leading to more robust outcomes.

Quick Start & Requirements

Installation is straightforward via npm:

• CLI: npm install -g @onevcat/argue-cli
• Library: npm install @onevcat/argue

Configuration involves initializing a config file (argue config init) and adding AI model providers (e.g., Claude, Codex via CLI or API). Debates are run using argue run --task "...". The system requires Node.js and npm. It supports various provider types: cli, api, sdk, and mock.

Highlighted Details

• Harness-Agnostic: Integrates with existing agentic workflows as a skill or library.
• Comprehensive Output: Generates detailed events.jsonl, result.json, and summary.md files locally.
• Interactive Viewer: Results can be visualized round-by-round via the argue view command or a hosted viewer at https://argue.onev.cat/.
• Action Execution: The representative agent can be configured to perform actions based on consensus, such as opening pull requests.
• Live Demo: A real-time example run is available at https://argue.onev.cat/example.

Maintenance & Community

The README does not provide specific details on notable contributors, sponsorships, or community channels like Discord or Slack.

Licensing & Compatibility

The project is licensed under the MIT license, which permits broad use, including commercial applications and linking within closed-source projects.

Limitations & Caveats

Reasoning passthrough for API providers is currently a no-op, meaning custom reasoning strings are accepted in configuration but not forwarded to the model. The behavior for handling insufficient participants has changed since version 0.5.0, requiring explicit configuration (--on-insufficient-participants fail) to restore the previous error-throwing behavior.