TL;DR
MULTILOOP introduces an automated framework for evolving the "harness"—the prompts, tools, and orchestration logic—of multi-agent LLM systems. While the approach successfully boosts accuracy on complex information-seeking tasks by 4.6%, the study reveals a critical bottleneck: automated optimizers excel at surface-level prompt tuning but struggle to implement the deep structural changes needed to fix fundamental orchestration failures.
What Problem Does This Paper Address?
Building effective LLM agents isn't just about the model; it’s about the harness—the surrounding code that defines how the agent uses tools, follows prompts, and interacts with other agents. Manually debugging these harnesses is becoming impossible as agent trajectories grow longer and more complex.
While recent research has automated harness optimization for single-agent systems, multi-agent systems present a unique challenge. Failures in these systems often cascade from poor orchestration—such as delegating tasks that are too broad, parallelizing steps that actually depend on each other, or failing to share the right context between agents. The authors set out to see if a coding agent could autonomously diagnose and repair these complex, multi-agent interaction failures.
Methods
The researchers developed MULTILOOP, a simple yet iterative optimization loop consisting of two main subroutines:
- EVALUATE: The current multi-agent system is run on a small training set (10 samples). The execution traces (rollouts) are converted into readable Markdown files that capture the orchestrator's thoughts, subagent steps, tool calls, and final answers.
- OPTIMIZE: A high-level coding agent (using the Antigravity SDK) acts as a "harness engineer." It reviews the rollouts, identifies failure patterns, and directly edits the harness source code. It has access to a history of past edits and scores to avoid regressing.
The system was initialized with a basic orchestrator-subagent architecture and optimized over 10 iterations using Gemini-3.1-Pro as the Optimizer and Gemini-3-Flash as the agent model.
Datasets
The study utilizes DeepSearchQA, a challenging benchmark designed for deep research agents.
- Training Set: 10 "hard" examples that the unoptimized system failed to solve.
- Test Set: 890 examples used to evaluate generalization.
- Task Type: Multi-step information seeking involving web search, PDF scraping, and Python execution in sandboxed environments.
Key Insights and Ideas
- Optimization Transfers, but Stays "Shallow": The optimizer successfully improved performance, but mostly through "low-hanging fruit." It added a new tool (
download_file) and refined prompts to be more specific about formatting and tool usage. However, it rarely touched the core logic of how agents communicate. - The Orchestration Bottleneck: The authors established a taxonomy of seven failure modes. As seen in the distribution below, while "Final Answer Composition" errors were significantly reduced, "Bad Orchestration" remained the most persistent and difficult category to automate away.
- Compute-Accuracy Trade-off: Optimization didn't just make the agents smarter; it made them "busier." The evolved harness spawned more subagents and allowed them to take more steps, effectively trading increased inference compute for higher accuracy.
- Prompt-Only vs. Full Harness Opt: A key finding was that allowing the optimizer to change the code (adding tools/logic) was superior to just changing prompts. Prompt-only optimization tended to overfit the training set, leading to longer but less effective agent trajectories.
Key Results
The multi-agent system, once optimized by MULTILOOP, achieved the highest accuracy on the DeepSearchQA benchmark among the tested configurations.
- Accuracy Boost: The optimized multi-agent harness reached 71.88% accuracy, a 4.6% improvement over the base version and 3.2% better than the best single-agent system.
- Trajectory Changes: The optimized system became more complex, increasing the average number of subagents spawned from 10.9 to 14.0 per question.
Limitations and Future Work
- Structural Inertia: Current coding-agent optimizers are hesitant to perform major architectural redesigns (e.g., changing the response schema between agents). Future work needs to find ways to encourage these "deep" structural changes.
- Small Training Sets: Due to the high cost of running complex agent trajectories (a single evaluation run can cost thousands of dollars), the training set was limited to 10 samples. This risks overfitting, as seen in the single-agent experiments where optimization actually degraded performance.
- Compute Matching: The authors note that the gains might partially stem from the system simply doing more work. Future studies should compare systems under a fixed "thinking token" or cost budget.
Summary updated