Three Minds Multi-Agent Collaboration Engine

Problem It Solves

A single agent reviewing code, generating ideas, or revising papers can become trapped by the same prompt context: it may return only conclusions, skip a second perspective, or let multiple agents override each other. Three Minds addresses a narrower problem: coordinating multiple role-based agents through a fixed process and detecting whether they have reached consensus.

How It Works

• Config-driven setup: Selects configs/code-review.json, configs/idea-brainstorm.json, or configs/paper-writing.json based on trigger keywords; falls back to the default collaboration config.
• Serial multi-round collaboration: Launches one persona subagent per round in general-purpose mode, avoiding inheritance of the main agent context. Later agents can see the key changes made earlier, reducing erratic perspective shifts.
• Consensus voting: Checks for [CONSENSUS: YES] in each agent's final reply. All YES ends the run; any NO or missing marker continues.
• Audit trail: Writes three-minds-{timestamp}.md with concise agent content, preventing full transcripts from bloating the record.

Boundaries and Notes

It fits tasks that benefit from multi-perspective convergence, such as code review, security review, brainstorming, and paper revision. It is not a substitute for tools that require real-time external actions, and consensus is not guaranteed: the run stops after maxRounds and outputs a record. Treat configs/ as read-only, and early termination is allowed if two consecutive rounds produce identical votes without substantive change.