AI

Dynamic Harness Search: Building Multi-Agent Systems Per-Query via Prediction

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 05:18:24 UTC.

Briefings

The SHIFT framework enables multi-agent systems to optimize harness selection per query using a local LLM that predicts utility balancing accuracy and execution cost. By replacing per-query search with a policy learned from Monte Carlo tree search, SHIFT achieves the highest mean accuracy of 80% across 9,193 tasks in six benchmarks. It outperforms 17 baselines by up to 7.2 percentage points and offers a cheaper mode using 32% fewer tokens while maintaining superior accuracy.

Key points

• SHIFT uses a local LLM to learn a policy for harness-building actions from search, predicting utility balancing accuracy and execution cost.
• Monte Carlo tree search constructs the optimal harness for each query using the SHIFT predictions, achieving 80% mean accuracy across 9,193 tasks.

Context

This research addresses the challenge of tailoring agent harnesses to specific queries, which previously required either expensive execution or manual design. SHIFT introduces a local LLM architecture that learns from search to balance accuracy and cost, significantly improving performance over existing prompting and workflow search methods.

Limitations

Results are based on 9,193 tasks across six benchmarks using a Gemini 3.5 Flash executor. The study does not provide causal claims about the generalizability of the policy or the specific reasons for the 7.2 percentage point improvement over baselines.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 18/10/2026, 05:06:29 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeifeumdm7l6cbbtjpqothjahicq3z5iyeubgaseafgck5lap6vbpcq