Your five-persona agent crew performs no better than one agent — how do you diagnose it?

## The symptom and what it usually means A crew of five carefully written personas that scores the same as one agent is not an unlucky result — it is the default result when the roles differ only in prose. This is worth stating plainly because the intuition runs the other way: it feels as though giving a model a distinct identity, goal and backstory should sharpen its behaviour, and frameworks that expose exactly those fields reinforce the feeling. The empirical picture is less flattering. When tool access and visible context are held constant, changes to persona wording move outcomes by small amounts on most tasks. The levers that move outcomes substantially are the ones that change *what the agent sees* and *what the agent can do*. ## Lever one: context isolation Model quality degrades as the context window fills with material — the standard term is context rot. An agent carrying twenty tool results, three abandoned hypotheses and a long transcript reasons worse about the next step than the same model given only what that step requires. That is the real mechanism behind subagent specialization. A role that receives a narrow brief, works in its own window, and returns a short summary is not smarter because it was told it is a specialist; it is more accurate because its window contains a high ratio of relevant to irrelevant tokens. The same effect explains why a reviewer role with a deliberately clean context catches defects the producing agent missed: the producer's window is saturated with the reasoning that led to the defect, and that reasoning is exactly what makes the defect look correct. So the diagnostic question is: does each of my five agents see a *different, smaller* slice of the problem? If they all see the same full history, I have five copies of one degraded context. ## Lever two: tool scoping The second lever is the action space. Tool-selection accuracy falls as the number of available tools rises, so an agent holding four relevant tools makes better choices than the same model holding thirty. Scoping tools per role therefore improves behaviour directly, independently of anything written in the prompt — and it does so in a way the model cannot override, because it is enforced outside the model. The combination is what specialization actually is: a smaller window and a smaller action space, with prose serving mainly to set priorities and output shape within those bounds. ## How to run the diagnosis Ablate, one variable at a time, against a fixed eval set. 1. **Baseline.** One agent, full tools, doing the whole task. Record score, tokens and wall-clock time. People skip this step and then cannot tell whether the crew helped. 2. **Persona ablation.** Keep the five agents, their identical tools and identical context, and replace all five personas with one neutral instruction. If the score barely moves, the personas were decoration. 3. **Scope ablation.** Restore the personas but give each role only the tools its job needs. Re-measure. Movement here tells you the action space was the problem. 4. **Isolation ablation.** Give each role only its own slice of context, returning a compressed result to the orchestrator. Re-measure. This is usually where the largest change appears. 5. **Partition check.** Ask whether any two roles could swap tasks without noticing. If yes, they are not specialized; they are duplicated. Run each configuration several times, because agent runs are nondeterministic and a single sample will happily confirm whichever hypothesis you started with. ## What to do with the answer If the ablations show the gain lives in isolation and scoping, redesign around those: fewer roles, each with a genuinely narrow window and a genuinely narrow tool set, with prose reduced to a job statement and a return contract. If the ablations show no configuration beats the single-agent baseline at equal token budget, that is a real and publishable result for your workload — 2026 papers report exactly this on a range of tasks — and the right move is to keep the single agent and spend the budget on better tools, better retrieval or a verification pass. A verification role is the one addition that tends to survive this analysis, because its value comes from being structurally uncontaminated rather than from being described as critical. ## The organisational reason this happens Worth naming, because a principal-level answer should. Persona text is the cheapest thing to change and the easiest to feel good about; it requires no plumbing, no permission model and no evaluation. Context partitioning and tool scoping require both engineering and an eval harness to prove they helped. Teams optimise the visible, cheap lever and then conclude multi-agent does not work. The fix is process as much as design: no role ships without a baseline comparison, and no crew grows a sixth agent without an ablation showing the fifth earned its place.