Multi-Agent Swarms

DevDiff v1.0.1 introduces local Multi-Agent Collaboration. Instead of relying on a single generalist LLM, DevDiff spawns a team of highly specialized local AI agents to analyze changes and build consensus.

Specialized Agents

• Architect Agent (ollama://llama3.1:8b): Analyzes code changes for architectural impacts, patterns, and structure.
• Security Auditor (ollama://codellama:13b): Scans for security vulnerabilities, secrets leakage, and CWE matches.
• Performance Engineer (ollama://llama3.2:3b): Monitors performance complexity and estimates bundle impact.
• Documentation Writer (ollama://llama3.2:3b): Drafts changelog files and guides.

Collaborative Process

1. Independent Analysis: Each agent reviews the diff and produces independent findings.
2. Discussion: Agents exchange messages to review and challenge each other's findings.
3. Consensus Building: Findings backed by more than 50% of the swarm are accepted into consensus.
4. Final Report: The documentation writer agent synthesizes the consensus findings into a final report.