Phoenix
Multi-agent GitHub issue resolution with baseline-aware safety checks
- Status
- Preprint
- My role
- Co-author & Architect
- Period
- January 2026 - Present
Overview
Phoenix coordinates six specialized LLM agents to resolve GitHub issues from triage through pull-request creation, checking proposed changes against a baseline test run.
The evaluation includes a curated SWE-bench Lite slice and a pilot across real repositories. These bounded results should not be interpreted as full-benchmark performance or a guarantee of correctness on arbitrary codebases.
The preprint is available as 'Phoenix: Safe GitHub Issue Resolution via Multi-Agent LLMs', co-authored with Kipngeno Koech, Muhammad Adam, and Joao Barros.
Methods and Tools
- LLM Agents
- Static Analysis Gates
- Dynamic Validation
- Python