Splitting a job across several agents costs reliability, and it earns that cost when the split is parallel work, permissions or tools rather than job titles.
- Errors multiply along a chain. Agents that are each usually right are, in sequence, right less often than any one of them, before counting what the handoffs lose.
- Each handoff is a lossy summary. The agent passing work on decides what matters without knowing what the next one needs.
- Failures get harder to trace. More agents means more traces and handoffs to search, and the error may have started well upstream of where it shows.
- Role names are not capabilities. Calling an agent a researcher changes the tone of its output, not what it can do.
- Try one agent first. If one agent with a longer brief could do the job, use that; split for parallel work, an enforced permission boundary, different tools, or checking against a fixed rubric.
A "multi-agent" setup means splitting a job across several AIs that pass work to each other — one researches, one writes, one checks.
It looks efficient. The problem is that each handover loses a little, and the losses stack. Multiply it out and three AIs that are each right 9 times in 10 are, together, right closer to 7 times in 10.
Multi-agent setups look impressive in a demo and, by the arithmetic below, are fragile in production. That is reasoning, not a measurement, and the useful question is when the trade is worth it.
Why it degrades
Errors compound across handoffs
If each agent is 90% reliable, three in sequence are not 90% reliable. Arithmetically, they are roughly 73% — and that is before the handoffs themselves introduce loss.Reasoning, September 2026 — arithmetic, not a measurement: 0.9 × 0.9 × 0.9 = 0.729. It holds only if each agent fails independently of the others; real chains can do better (a later step catches an earlier error) or worse (errors feed each other). No study of agent chains is cited for this figure.
The arithmetic is unforgiving in a way a demo never shows, because a demo runs once and succeeds.
Each handoff is a lossy summary
Agent A passes agent B a summary of what it did, not what it saw. The nuance that mattered is exactly what gets compressed out, because summarising means deciding what is unimportant — and A does not know what B needs.
Failures become unattributable
A single agent that fails leaves one trace. Five agents leave five, plus the handoffs between them, and the wrong answer may have originated three steps upstream of where it became visible. See diagnosing a run — it gets harder as the number of agents grows.
Job titles are not capability
Calling one agent “researcher” and another “editor” changes the tone of their output, not the tools or access either one has, and it does not reliably make either one better at the job. The role names are a convenience for you, not a capability boundary.Zheng, Pei, Logeswaran, Lee & Jurgens, When “A Helpful Assistant” Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models, Findings of EMNLP 2024, read at source 23 Sep 2026: across 162 roles, 4 model families and 2,410 factual questions, “adding personas in system prompts does not improve model performance across a range of questions compared to the control setting where no persona is added.” The same paper finds a persona can move accuracy on individual questions, but “the effect of each persona can be largely random.”
When it genuinely helps
- Genuinely parallel work — twenty independent items processed at once. No handoffs, so no compounding.
- A real permission boundary — one agent reads production, another writes to staging, and the split is enforced rather than described. That is a permission tier, not a persona.
- Different tools, not different personalities — one has web access, one has the database. The separation is capability, so it means something.
- Adversarial review — one produces, one checks against a fixed rubric. This can work because checking is easier than generating: a formal result puts a model’s error rate when generating at a floor of roughly twice its error rate when judging whether an output is valid. That floor says the gap exists; it does not measure its size.Kalai, Nachum, Vempala & Zhang, Why Language Models Hallucinate, OpenAI, Sep 2025, §1.1 and Corollary 1, read at source 17 Sep 2026: “(generative error rate) ≳ 2·(IIV misclassification rate)”. A formal lower bound, not an empirical measurement. Corrected 17 Sep 2026: this line put a size on the gap (about two times), cited Theorem 3, and quoted a sentence that is not in the paper.
The test before you build one
Could one agent do this with a longer brief? If yes, use one agent and a longer brief.
Multi-agent earns its complexity when the split is parallelism, permissions, or tools. It does not earn it when the split is job titles.