SwarmResearch orchestrates coding agents to beat SOTA on 13/15 tasks
New agent harness avoids single-approach dead ends with a shepherd and swarm
Long-running coding agents like AutoResearch can persistently optimize open-ended problems, but they often converge on one strategy and then only make low-level edits, missing superior alternatives. The authors identify two root causes: accumulating all context in a single agent and exposing only one program state for editing. To solve this, they built SwarmResearch—a harness where a central Shepherd Agent maintains a global view of the search landscape and a population of Search Agents each work independently on their own git branch, exploring different approaches in parallel.
In tests across 15 open-ended optimization tasks, SwarmResearch outperformed state-of-the-art LLM-guided evolution and multi-agent baselines on 13 tasks. The key advantage came from higher-level exploration that avoided local optima. Additionally, SwarmResearch's orchestrator-guided scaling adapts parallelism dynamically based on search depth, yielding better results than simply adding more serial or parallel agents. This work demonstrates that careful orchestration of multiple independent agents can unlock more creative solutions in code optimization and scientific discovery.
- Shepherd Agent uses global context to coordinate multiple Search Agents, each working in isolated git branches to prevent convergence on one approach
- SwarmResearch beats LLM-guided evolution and multi-agent methods on 13 out of 15 open-ended optimization tasks
- Dynamic parallelism adaptation outperforms fixed scaling of serial or parallel agents by allocating resources based on search depth
Why It Matters
Enables more creative, high-level code optimization in AI research, potentially accelerating automated scientific discovery.