competitive-agents

Two agents with complementary philosophies generate competing Claude Code plugin implementations and cross-review each other's work. A built-in judge then selects the winner, with an optional fusion phase to combine strengths from both versions.

A phased workflow where two agents generate competing plugin implementations, cross-review each other's code, and a judge selects the winner. It includes v1/v2 generation cycles, optional fusion of strengths, and ADR generation for architectural decisions.

Describe the plugin you want to build and trigger the competition with a phrase like compete to start the two-agent generation. This initiates the phased workflow including parallel generation, cross-review, and final judge decision.

Yes, after the judge selects a winner, an optional fusion phase combines the strengths of both competing agent implementations into a single plugin. This ensures the final output retains the best elements from both versions.

Yes, the workflow generates ADRs (Architecture Decision Records) to document architecture decisions made during plugin generation. It also supports planning-interview contextual data and uses SKILL.md-driven prompts to guide the phased workflow.

The competitive agent workflow requires more processing overhead than single-agent generation due to coordinating parallel workstreams, cross-review, and judge phases. It is best suited for complex plugin tasks where quality gains justify the additional cycles.