When Multi-Agent Systems Help—or Hurt

Single agent, Multiple sub-agents, and Agent society are three points on a scale of coordination cost. The first is the default. Move off it only when each agent has a distinct capability, trust boundary, or parallel artifact — when the split buys something the single agent could not do. Without that, the extra agents mostly agree with each other, and the agreement is mistaken for a check.

The rest of this lesson is the evidence for that ordering. Every claim in it comes from a published measurement: one vendor's engineering report, and peer-reviewed papers with the numbers in them. In every one of them, the multi-agent design had to be measured against a single-agent baseline before anyone could say whether the split had paid.

A multi-agent system coordinates several model-powered components that may have different roles, tools, contexts, owners, or trust domains. It can improve parallel work, specialization, isolation, and independent review.

Additional agents also multiply communication, state, cost, latency, failure paths, and correlated error. Role names do not create independent expertise or evidence. One of those failure paths has been measured and named: problem drift. It was quantified across ten tasks. Drift appeared in 76–89% of generative tasks, against 7–21% of high-complexity tasks. Eight human experts then analysed 170 drifting debates. Becker and five colleagues report what they found: “We find the most common issues related to this drift are the lack of progress (35% of cases), low-quality feedback (26% of cases), and a lack of clarity (25% of cases).” Their DRIFTPolicy mitigated 31% of cases. That is the honest measure of how much a mitigation bolted onto the architecture can win back.

Read the causes. None of them is a shortage of expertise. Lack of progress, low-quality feedback and lack of clarity are costs of the conversation itself. They arrive with the second agent. They grow with every turn after it.

That design has been run as an experiment. A 2025 study of multi-agent debate held the agents fixed and varied only the decision protocol, across seven protocols. Voting protocols improved reasoning tasks by 13.2%. Consensus protocols improved knowledge tasks by 2.8%. Two methods that deliberately raised answer diversity added up to 3.3% (All-Agents Drafting) and up to 7.4% (Collective Improvement).

The finding that indicts the six-specialist design is the shape of the curve. Kaesberg and Becker state it flatly with their co-authors: “Increasing the number of agents improves performance, while more discussion rounds before voting reduce it.” The gains came from the aggregation rule and from raising answer diversity. The discussion between similar agents is the part that produced the conflicting conclusions and the token bill. It is also the part measured to make things worse the more of it you buy.

Adding an agent is justified only when Separable work, Distinct capability, Coordination mechanism, and Aggregation all have answers. Aggregation and the Overhead test should belong to separate owners, each with its own test.

The Overhead test — ablation against one-agent and deterministic baselines — is the one teams skip. There is a published account of what skipping it hides. Three published agent architectures, Reflexion, LDB and LATS, were re-run on the 164 problems of the HumanEval benchmark, five runs each. Kapoor and four colleagues compared them against a trivial single-model baseline: retry, with the temperature gradually raised from 0 to 0.5. Their conclusion is one sentence: “There is no significant accuracy difference between our warming strategy and the best-performing agent architecture.” The prices differ enormously. Reflexion and LDB cost over 50% more than that baseline. LATS cost over 50 times more.

That is the whole argument for the ablation. Without a baseline run at the same task, an architecture that costs 50 times more and is no more accurate looks exactly like an architecture that works.

Every additional agent should have to justify itself in review, and the review is worth running on a workflow that already has several. A review that lets every agent through has not really been a review. It has only confirmed what the design already assumed.

The two steps that do the work are Build the single-agent baseline — same model, tools, and total budget — and Require measured gain. A benchmark presented at ICML in 2024 shows what those steps catch. It measured debating and prompting strategies against cost, time and accuracy. Its finding: “Importantly, we find that multi-agent debating systems, in their current form, do not reliably outperform other proposed prompting strategies, such as self-consistency and ensembling using multiple reasoning paths.” The debate architecture did not reliably beat the cheaper single-agent protocols it was built to improve on. Only a comparison against those protocols, at matched cost and time, could have revealed it.

So the review has an order. Name the bottleneck that one agent cannot address well. Define what each agent produces and how it will be checked. Estimate the coordination cost in messages, duplicated context, merge work and new failure modes. Build the single-agent baseline at the same total budget. Keep the multi-agent design only where it improves verified outcomes or containment. A design that cannot clear its own baseline has bought coordination cost and nothing else.

The strongest multi-agent design often uses fewer agents with clearer artifacts, permissions, and stop rules.

Before adding the third agent, ask what it can do that the existing agents cannot. If the answer is a capability, a permission boundary, or an artifact the others do not produce, the addition is justified. If the answer is that it will double-check the others, the literature already ran that comparison. A 2024 ACL paper asked whether multi-agent discussion is really the key to LLM reasoning, and re-tested the claim under systematic experiments. Wang and four colleagues answer it in their abstract: “Interestingly, our results show that a single-agent LLM with strong prompts can achieve almost the same best performance as the best existing discussion approach on a wide range of reasoning tasks and backbone LLMs.” Multi-agent discussion came out ahead only where the prompt contained no demonstration — only where the single agent had been handicapped first.

That is the test to apply to the third agent. Strengthen the prompt of the agent you already have. Then ask whether the addition still wins. If the discussion only beats a single agent you deliberately weakened, expect agreement rather than scrutiny, and keep the agent count where it is.