A coding agent can tune AI agent teams, but their teamwork flaws mostly remain

The wrapper around the model

When people talk about an AI agent, they usually talk about the model inside it. This paper starts from a different point. As arxiv.org shows in the abstract, how well an agent performs depends on the model and also on the harness: the prompts, tools and rules for coordination wrapped around it.

Think of the model as a talented employee and the harness as the job description, the toolbox and the office procedures. The same employee can do very different work depending on those things. Tuning the harness is therefore a real way to improve results without touching the model at all.

Letting a coding agent do the tuning

Researchers have recently begun to automate that tuning. A coding agent reads the execution traces, which are records of what the agent did step by step. It works out where things went wrong and then edits the harness code to fix it. According to the paper, that earlier work focused on single-agent systems.

Multi-agent systems are harder. Here, a fix may not be about what one agent does. It may require changing how the agents interact, such as who gets which part of the job and what they tell each other.

The authors ask whether automated tuning can diagnose and repair failures in these multi-agent setups. They test this through a case study on complex information-seeking tasks from a benchmark called DeepSearchQA (Gupta et al., 2026). These are questions where an agent must dig up an answer rather than recall it.

What MULTILOOP did and the results

The method, called MULTILOOP, is deliberately plain. A coding agent repeatedly goes through three steps: it analyzes how the system failed, evaluates the result, and edits the harness. Then the cycle starts again.

The data it needed was small. With only 10 training samples, the evolved harness improved accuracy by 4.6% over the base harness. It also beat the best single-agent system by 3.2%. So a team of agents, once tuned this way, edged out the strongest solo setup the researchers compared against.

• Base harness: beaten by 4.6%.
• Best single-agent system: beaten by 3.2%.
• Training data used: 10 samples.

These are modest gains on one family of tasks, and the source reports them as a case study, not a general rule. It is also an early arXiv posting, which means other researchers have not yet reviewed it.

The catch: teamwork failures stayed

The more interesting finding is about what did not improve. The researchers built a taxonomy, meaning a sorted catalog of failure types, and applied it both before and after optimization. The failures that define multi-agent systems, which the authors call orchestration failures, largely persisted.

In plain terms, the tuning loop made the system somewhat more accurate, but it did not repair the way the agents coordinate. The automated editor seemingly found improvements that were easier to reach, not the structural ones.

That is why the authors argue for a different kind of optimization. They want methods that push toward structural changes in how agents delegate tasks and share information, instead of only polishing prompts and tool settings.

What to take from it

For builders, the message is that a cheap automated loop is worth trying and can pay off with very little data. It is not a cure for multi-agent design problems, and someone still has to think about how the agents divide the work.

For everyone else, it is a useful reminder that an AI answer reflects more than the model's raw ability. The instructions and structure around it matter too. We only have the abstract here, so details such as which models were used or how the failure categories break down are not available to us.