Dynamic Harness Search: Building Multi-Agent Systems Per-Query via Prediction
SHIFT predicts the right agent setup for each query instead of running a live search over alternatives.
The paper frames an “agent harness” as the mix of roles, instructions, tools, and communication structure used to solve a task. Its system trains a local LLM architect and value function from measured executions, then uses Monte Carlo tree search to build a harness per query. Across 9,193 tasks on six benchmarks, SHIFT reports about 80% mean accuracy with a Gemini 3.5 Flash executor, beating 17 baselines. A cheaper mode also beats every baseline while using 32% fewer execution tokens than the strongest one. Source: HF Daily Papers' note
The paper frames an “agent harness” as the mix of roles, instructions, tools, and communication structure used to solve a task. Its system trains a local LLM architect and value function from measured executions, then uses Monte Carlo tree search to build a harness per query. Across 9,193 tasks on six benchmarks, SHIFT reports about 80% mean accuracy with a Gemini 3.5 Flash executor, beating 17 baselines. A cheaper mode also beats every baseline while using 32% fewer execution tokens than the strongest one. Source: HF Daily Papers' note
score 5