SHIFT: Dynamic Multi-Agent Harness Search Per-Query via Predictive MCTS Outperforms Static Baselines by 7.2%
Designing optimal multi-agent harnesses—specifying roles, prompts, tool registries, and communication graphs—is foundational to task performance. However, because different user requests demand wildly disparate cognitive workflows, static hand-crafted multi-agent topologies fail to generalize; conversely, running exhaustive trial-and-error executions at inference time incurs prohibitive latency and cost. Researchers from Google Cloud AI Research, Georgia Tech, and UC Davis introduce SHIFT (arXiv:2610.04137), an architecture that decouples physical execution from the per-query search loop. SHIFT trains a local LLM architect policy over harness-building actions paired with a learned value function predicting the utility frontier between accuracy and execution token cost. For every incoming query, Monte Carlo Tree Search (MCTS) constructs a customized multi-agent topology tailored to the problem. Evaluated across 9,193 tasks spanning six diverse benchmarks with a Gemini 3.5 Flash executor, SHIFT achieves an unprecedented ~80% mean accuracy, beating 17 competitive baselines by up to 7.2 percentage points while slashing execution tokens by 32%.
Key Takeaways
- Introduces SHIFT, a framework dynamically generating tailored multi-agent harnesses per-query via predictive Monte Carlo Tree Search
- Decouples execution from the search loop using an LLM architect policy and a learned value function balancing accuracy against token cost
- Achieves ~80% mean accuracy across 9,193 tasks in six benchmarks with Gemini 3.5 Flash, outperforming 17 baselines while cutting execution tokens by 32%