MULTILOOP: Evolving Multi-Agent Harnesses for Complex Information-Seeking Tasks
Daniela Gottesman, Gal Yona, Mor Geva, Roee Aharoni and colleagues at Google Research and Tel Aviv University test whether a coding agent that edits harness code can repair a multi-agent research system, and find that it improves accuracy without fixing the orchestration failures specific to multi-agent setups.
Ask this paper
Setup. MULTILOOP is a plain optimization loop: a coding agent reads execution traces from a multi-agent harness on DeepSearchQA, diagnoses failures, edits the harness code and re-evaluates.
Result. With only 10 training samples, the evolved harness is 4.6% more accurate than the base multi-agent harness and 3.2% more accurate than the best single-agent system.
Where the gain comes from. The improvements are prompt refinements and one new tool. The optimizer did not change how agents delegate tasks or what information they pass to each other.
Failure taxonomy. A before-and-after failure analysis shows that orchestration failures, such as poor task decomposition and lost context between agents, largely persist after optimization.
Takeaway. Harness optimizers that work for single agents need explicit pressure toward structural changes in delegation and information sharing before they can fix multi-agent systems.
Abstract
The performance of LLM-based agents depends not only on the underlying model but also on the harness: the prompts, tools, and orchestration logic around it. Recent works automate harness optimization with coding agents that analyze execution traces, diagnose failures, and edit the harness code; however, those focused on single-agent systems. Multi-agent harnesses are harder to optimize as fixing them requires changing how agents interact in addition to how each behaves. We study whether automated harness optimization can diagnose and repair failures in multi-agent systems through a case study on complex information-seeking tasks from DeepSearchQA [Gupta et al., 2026]. We apply MULTILOOP, a simple optimization loop in which a coding agent iteratively analyzes, evaluates and edits the harness. With only 10 training samples, the evolved harness improves accuracy by 4.6% over the base harness and by 3.2% over the best single-agent system. However, a taxonomy of failure modes before and after optimization shows that the orchestration failures that characterize multi-agent systems largely persist. Our findings motivate harness optimization methods that encourage structural changes to how agents delegate tasks and share information.