- In short
- Narrow decomposition failure is a multi-agent fault where the coordinator splits a task too narrowly, so important subtopics are never assigned and the final output is incomplete. The root cause is always the coordinator's decomposition, not the subagents that executed their assigned slices correctly.
What narrow decomposition failure is
Narrow decomposition failure is a precise diagnosis the exam wants you to make on sight. A multi-agent system returns an answer that reads well and is internally consistent, but it quietly omits whole parts of what was asked. When you investigate, every subagent did its assigned job correctly. The fault is upstream: the coordinator carved the task too narrowly, so the missing material was never assigned to anyone. No spoke can cover a subtopic it was never given.
This knowledge point is rated at the evaluate level because the skill is causal attribution under a misleading surface. The visible work was done by subagents, so the naive conclusion is that a subagent fell short. The architect's job is to resist that and trace the gap to its real source: the coordinator's decomposition, a coordinator responsibility.
The reason this is genuinely hard, and therefore worth testing, is that human attention follows the visible artefacts. When you read a report, you judge the sections that are there; you do not naturally notice the sections that are absent, because absence leaves no trace on the page. Narrow decomposition failure exploits exactly that blind spot. Diagnosing it well means deliberately checking the answer against the full scope of the original request rather than against itself, asking "what should be here that isn't" instead of "is what's here good." That shift in question is the whole skill.
- Narrow decomposition failure
- A multi-agent fault where the coordinator partitions a task too narrowly, leaving important subtopics unassigned, so the final output is incomplete despite every subagent executing its slice correctly.
The canonical example
The classic illustration the exam draws on is a research task framed broadly, say, "the impact of AI on creative industries", that the coordinator decomposes into only "visual arts." The visual-arts subagent does excellent work. But music, writing, and film were never assigned, so the briefing silently omits three-quarters of the field. The output is polished and wrong in the way that matters most: it is incomplete, and nothing in it signals the absence.
What makes this dangerous is that the gap is invisible from the inside. The synthesis reads as a finished answer because it is internally coherent; it simply has no awareness of what it never covered. A reader who does not already know the field may not notice. This is why decomposition quality is a first-class design concern rather than an afterthought.
Why the coordinator, never the subagent
The attribution rule is absolute and worth stating starkly: in a coverage failure, the root cause is the coordinator, not the spokes. The logic is mechanical. A subagent operates only on the slice it was handed, in its own isolated context. It cannot research a topic it was never assigned, and it has no view of the whole task to notice the omission. So a missing subtopic can only mean the coordinator failed to assign it.
This is the inverse of blaming the messenger. The subagents are the messengers, faithfully reporting on their slices; the editor who decided which slices to commission is the coordinator. When you evaluate a multi-agent failure on the exam, the presence of a coverage gap should point your attention immediately and entirely at the decomposition step.
How to spot it: the diagnostic signature
Narrow decomposition failure has a recognisable fingerprint, and learning to read it is the core skill. The tell is the combination of high local quality and missing scope: every section that exists is accurate and well-argued, yet the whole answer omits parts of the question. When you see competent work plus an absent topic, your attention should jump immediately to how the task was divided, not to how well any piece was executed.
This signature is what distinguishes a decomposition failure from other faults. If sections were present but wrong, you would suspect a subagent's reasoning or its inputs. If the answer were empty, you would suspect context passing. But coherent, correct sections with a hole in coverage can only mean the hole was never assigned. Training yourself to read that pattern quickly is what lets you answer these evaluative questions with confidence rather than getting drawn into critiquing the visible, blameless work.
The relationship to iterative refinement
Narrow decomposition failure and iterative refinement are two defences against the same enemy: incomplete coverage. Good decomposition is the preventive measure, split the task well up front so nothing is left out. Iterative refinement is the corrective measure, after synthesis, inspect the draft for gaps and re-delegate to fill them. A robust coordinator uses both, because even careful decomposition occasionally misses something, and a refinement pass is the safety net that catches it.
Seeing the two as complementary clarifies why neither alone is sufficient. A coordinator that decomposes perfectly but never reviews its draft will still ship the rare gap that slipped through. A coordinator that relies entirely on refinement to patch a chronically narrow decomposition will loop endlessly chasing omissions it could have prevented. The exam rewards the architect who reaches for the right remedy: fix the decomposition for systematic gaps, add refinement for the residual ones.
The fix, and the fix that doesn't work
Because the cause is the decomposition, the fix is to broaden the decomposition, make the coordinator partition the goal so that every important subtopic is assigned to some subagent. In practice that often means improving the coordinator's prompt or strategy so it enumerates the full scope before delegating, rather than latching onto the first facet.
The tempting non-fix is to throw more subagents at the problem. But adding subagents does nothing if the decomposition still fails to hand them the missing topics; you just have more idle or redundant spokes. Likewise, sharpening a single subagent's prompt cannot help, because that subagent was never the bottleneck. The lever is always the coordinator's partitioning logic. Fix where the work is divided, and the coverage follows.
In practice, improving the decomposition usually means changing how the coordinator is instructed to approach the goal. A coordinator told to "research this topic" may latch onto the first facet it thinks of; one told to "enumerate the major dimensions of this topic, then assign a subagent to each" is far more likely to span the full scope before delegating. The remedy lives in the coordinator's strategy and prompt, because that is where the partitioning decision is actually made. This is a useful tell for the exam: when the correct fix is phrased as changing the coordinator's planning or decomposition prompt, and the distractors all act on subagents, the coordinator-side answer is almost always right.
Worked example: a market-research briefing with a hole
Worked example
A coordinator is asked to assess a product's competitive landscape, and the resulting report only discusses direct competitors.
The brief was "assess the competitive landscape," which spans direct competitors, indirect substitutes, and emerging entrants. The coordinator, however, decomposed it into a single slice, "list and compare direct competitors", and dispatched one subagent. That subagent produced a sharp, accurate comparison of direct rivals. The final report is confident and well-written, and it entirely ignores substitutes and new entrants.
A reviewer's first instinct is to ask the subagent to "be more thorough." That cannot help; the subagent thoroughly covered exactly what it was told to. Tracing the gap correctly, the architect sees that substitutes and entrants were never assigned, a narrow decomposition failure. The fix is to broaden the coordinator's decomposition so the landscape is split into all three categories, each routed to a subagent. Only then does the report become complete, and the change happened at the hub, not the spoke.
Misconceptions to discard
Misconception
If the report misses a major topic, the subagent that wrote it must have done a poor job.
What's actually true
Misconception
Coverage gaps are fixed by adding more subagents to the system.
What's actually true
Decomposition quality versus subagent quality
A clarifying way to hold this knowledge point is to separate two independent axes of multi-agent quality. One axis is decomposition quality: did the coordinator carve the task so that every important part was assigned? The other is execution quality: did each subagent do its assigned slice well? Narrow decomposition failure is a defect on the first axis that masquerades as a problem on the second, because the visible artefacts, the subagent sections, all look fine.
Keeping the axes distinct stops you from misdiagnosing. High execution quality can coexist with low decomposition quality, and that combination is exactly the incomplete-but-polished report. No amount of lifting execution quality, better subagent prompts, stronger models on the spokes, moves the decomposition axis. The only lever for coverage is the partitioning itself. An architect who instinctively asks "is this an execution problem or a decomposition problem?" will route every coverage question to the right fix on the first try.
The other extreme: over-decomposition
Narrow decomposition usually means under-coverage, but splitting a task too finely is its own failure mode, and a complete picture recognises both ends of the spectrum. When a coordinator carves a goal into too many tiny, over-isolated subtasks, the subagents can lose sight of the larger objective, duplicate each other's work, or return fragments that are hard to recombine. The coordination overhead of briefing and merging a swarm of micro-subagents can outweigh any parallel speed-up, so more decomposition stops helping and starts hurting. Anthropic's guidance is to scale effort to query complexity rather than fragment by default: simple fact-finding may need only a single subagent, while genuinely complex research is what warrants many with clearly divided responsibilities.
A related hazard is the "game of telephone." If every intermediate finding must be relayed back through the coordinator and re-summarised before the next step, detail erodes at each hop and global constraints get dropped. Anthropic's mitigation is durable handoffs: have the lead agent write its plan to external memory before spawning work, and let subagents write their outputs to a shared filesystem or artifact store instead of funnelling everything through the coordinator's context. Pairing sensible decomposition with external memory keeps each subagent anchored to the wider goal, which is the same continuity that guards against both the coverage gaps this knowledge point centres on and the fragmentation of its opposite.
How this is tested on the exam
Task 1.2 will present a multi-agent output that is coherent but incomplete and offer four candidate causes or fixes. The distractors point at the subagents, "the research subagent was not thorough," "add a second subagent," "improve the subagent's prompt." The correct answer points at the coordinator's decomposition. The evaluative skill being tested is your ability to look past the visible, competent work of the spokes and locate the silent omission at the hub. Because this knowledge point depends on a firm grasp of coordinator responsibilities, it is effectively the diagnostic capstone of decomposition: see the gap, name the cause, broaden the split.
A research coordinator is asked to summarise 'the effects of remote work,' but the summary covers only productivity, omitting wellbeing, culture, and cost. Each subagent's section is accurate. What is the correct diagnosis and fix?
People also ask
Why does my multi-agent system miss topics?
Who is responsible for incomplete coverage in multi-agent systems?
How do you fix a coverage gap in a coordinator?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
Anthropic: How to Build Multi Agent Systems
Why watch: Walks through Anthropic's lead-agent-and-subagents architecture and how the coordinator decomposes a query into subtasks, showing why incomplete coverage traces back to the coordinator's decomposition rather than the subagents.
More videos for this concept
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.