- In short
- Coverage gaps in a multi-agent system's output trace back to the coordinator's initial decomposition, not to the subagents that executed it. A decomposition that names only a subset of the true subtopics leaves the rest uninvestigated by any subagent. Adding more subagents does not fix a narrow decomposition, the decomposition prompt itself must be widened, and reviewing decomposition quality means checking coverage against the full scope before subagents are dispatched.
The root cause of an incomplete answer
When a multi-agent system returns an answer that is missing whole areas it should have covered, the instinct is to blame the subagents, one of them must have done its job poorly. The Claude Certified Architect - Professional (CCAR-P) exam treats resisting that instinct as an apply-level skill because the real root cause is almost always upstream: the coordinator's decomposition never named the missing area, so no subagent was ever assigned to it. A subagent can only investigate what it was given; a subtopic the decomposition omitted is invisible to every subagent. The coverage gap is a decomposition defect wearing the costume of a subagent failure.
This reframes where you look. A subagent that returns a poor result for its assigned unit is an execution problem. A whole area absent from the output is a coordination problem, the decomposition was too narrow. Telling these apart is what lets you fix the right layer instead of the convenient one.
- Coordinator task decomposition quality
- The completeness of the coordinator's initial breakdown of a job into subtopics. Coverage gaps in a multi-agent system's output trace to this decomposition, not to the subagents that executed it, because a subtopic the coordinator never named is investigated by no subagent. Reviewing decomposition quality means checking coverage against the full scope before subagents are dispatched.
A narrow decomposition leaves scope uninvestigated
The mechanism is direct. The coordinator decomposes the request into a set of subtopics and dispatches one subagent per subtopic. If that set names only a subset of the request's true subtopics, the omitted ones are assigned to nobody, and the synthesis can only combine what was investigated. The output looks like a complete answer because every named subtopic was covered well, but it is silently missing the scope the decomposition never mentioned. A market analysis that decomposes into "competitors and pricing" but omits "regulatory environment" will produce a confident report with no regulatory coverage at all, and no subagent is at fault, because none was asked.
Adding subagents does not widen scope
The tempting fix, when an answer looks thin, is to throw more subagents at it. This fails, and seeing why is the point. More subagents run against the same, still-narrow decomposition, so they re-investigate the subtopics that were already named rather than reaching the ones that were omitted. You get more depth on the covered areas and still zero coverage of the missing ones. The scope of a multi-agent investigation is set by the decomposition, not by the subagent count, so the only fix for a coverage gap is to widen the decomposition prompt itself to include the omitted subtopics.
Review coverage before dispatch
Because the gap is set before any subagent runs, the place to catch it is before dispatch. Reviewing decomposition quality means checking the coordinator's breakdown against the full scope of the original request, does the decomposition name every subtopic the request genuinely requires, and widening it where it falls short, before subagents are launched. This is cheaper and more reliable than diagnosing a thin output after the fact, because at review time the omission is a missing line in a list rather than a hard-to-spot absence in a finished report.
What the CCAR-P exam trips candidates on
The exam tests two traps. The first is blaming a subagent for an incomplete answer when the coordinator never assigned it the missing subtopic. A scenario will point at a subagent's output and imply it underperformed, and the credited answer traces the gap to the decomposition, which never named the missing area, so no subagent could have covered it.
The second is fixing perceived coverage gaps by adding more subagents to an unchanged, still-narrow decomposition. A scenario will propose scaling up the subagent count, and the reliable reading notes that more subagents run against the same narrow decomposition and re-cover named subtopics without reaching the omitted ones, so the decomposition must be widened instead.
Worked example
A research orchestration is asked to produce a full competitive analysis of a market. The report covers competitors, pricing, and product features thoroughly but says nothing about the regulatory environment, which the brief clearly required. The team blames the subagents and proposes doubling their number. Diagnose the real cause and the right fix.
Trace the gap to its origin. The report is thorough on competitors, pricing, and features, which means the subagents assigned those subtopics executed well. Regulatory environment is entirely absent, not shallow, absent, which is the signature of a subtopic that was never assigned to any subagent. Look at the coordinator's decomposition: it named competitors, pricing, and features, but omitted the regulatory environment. No subagent covered it because none was asked to, so blaming the subagents is looking at the wrong layer.
The proposed fix, doubling the subagent count, does not address this. More subagents run against the same decomposition that still names only competitors, pricing, and features, so they add depth to those three areas and still produce zero regulatory coverage. The scope of the investigation is fixed by the decomposition, not by how many subagents execute it, which is why adding subagents cannot reach an unnamed subtopic.
The correct fix is to widen the decomposition prompt so it explicitly names the regulatory environment as a subtopic, then dispatch a subagent for it. More broadly, the review should have caught this before dispatch, by checking the coordinator's breakdown against the full scope of the brief, which listed the regulatory environment. Reviewing decomposition coverage up front turns a missing report section into a missing list item that is easy to spot and cheap to fix.
Common misreadings to avoid
Misconception
If the output is missing a whole area, one of the subagents failed to cover it.
What's actually true
Misconception
Adding more subagents will close a coverage gap.
What's actually true
How this shows up on the exam
Domain 1 questions on this knowledge point present an incomplete multi-agent output and ask for the root cause and fix. The reliable reading traces a missing area to the coordinator's decomposition rather than the subagents, rejects adding subagents to an unchanged decomposition, and prescribes widening the decomposition, ideally caught by reviewing coverage against the full request scope before dispatch.
Decomposition quality builds on orchestrator-subagent fan-out and fan-in, where the orchestrator owns decomposition, and connects to decomposition granularity and the sequencing concerns of dependency ordering in decomposition. Diagnosing which layer caused a gap also relies on observability across a multi-agent pipeline.
A multi-agent system tasked with a comprehensive security audit returns findings on network and application layers but nothing on physical security, which the scope required. Each subagent's assigned area was covered well. What is the correct diagnosis and fix?
People also ask
Why do coverage gaps trace to the coordinator, not the subagents?
Does adding more subagents fix an incomplete answer?
How do you review decomposition quality?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
No videos curated for this concept yet
We are still curating the best official and community videos for this topic.
Official prep for this domain
Anthropic's own free prep module for this part of the syllabus, on the official prep course. Free with an Anthropic Academy sign-in.
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.