The Coordinator Trap — Avoiding Scope Narrowing in Task Decomposition

⚡ Exam Tested 8 min +35 XP

💡

THE ANALOGY

A project manager who is asked to 'improve customer satisfaction' and only assigns work on the website UX — ignoring customer service, product quality, and pricing, which are all part of customer satisfaction. The work done is high quality. The scope was wrong.

⚠️ EXAM TRAP — The Wrong Answer People Choose

The exam presents a coordinator that has completed its assigned subtasks successfully, but the final output is incomplete because the decomposition was too narrow. The question asks for the root cause — it's always the coordinator's decomposition, not the subagents' execution.

KEY POINTS

1 The coordinator trap: a coordinator decomposes a broad task into subtasks that are internally consistent but collectively miss significant portions of the required scope.

2 Subagents executing coordinator trap subtasks will succeed — they do exactly what they're asked. The failure is in what they were asked.

3 Root cause identification: when the exam shows subagents that worked correctly but output is incomplete, look at the coordinator's decomposition first.

4 Prevention: the coordinator should explicitly enumerate the full scope before creating subtasks, then verify all scope is covered.

5 Scope enumeration is a separate step from decomposition — enumerate first, decompose second.

Why Coordinators Fall Into the Trap

Coordinators are doing two things when decomposing a task:

Understanding the full scope of what needs to be done
Breaking that scope into workable subtasks

The trap happens when step 1 is incomplete — the coordinator forms a coherent picture of ONE part of the scope and decomposes that, missing other parts entirely.

This is especially common with:

Broad categories: “creative industries,” “enterprise software,” “financial services”
Ambiguous scope: “analyze our product” — which products? which aspects?
Familiar domains: coordinators tend toward subtopics they’ve seen more examples of

Diagnosing the Trap in Exam Questions

The exam presents the trap as a symptoms-and-cause question. The symptoms are always:

Subagents completed successfully
Output quality within each subagent’s scope is high
But the aggregate output is missing entire categories

When you see this pattern, the root cause is always the coordinator’s decomposition — not:

Subagent tool selection (subagents used correct tools for their scope)
Subagent prompts (subagents followed their instructions correctly)
Synthesis logic (synthesis correctly combined what it received)
Model capability (model performed well within the given scope)

The diagnosis test: Ask “Did each subagent do what it was asked?” If yes → coordinator trap.

The Scope Enumeration Prevention Pattern

Add a scope verification step BEFORE decomposition:

coordinator_system_prompt = """
You coordinate research tasks using specialized subagents.

MANDATORY PROCESS — follow this exactly:

Step 1: SCOPE ENUMERATION
Before creating any subtask plan, explicitly list ALL categories and aspects
that fall under the research topic. Be comprehensive — not just what you know
well, but everything that belongs in scope.

Step 2: SCOPE VERIFICATION  
Review your enumerated list. Ask: "Is there anything that belongs to this topic
that I haven't listed?" Add any missing categories.

Step 3: COVERAGE CHECK
Create your subtask plan. For every item in your scope enumeration, verify
there is at least one subtask that covers it. If any enumerated item has no
subtask — add one before proceeding.

Step 4: SPAWN SUBAGENTS
Only after Step 3 is complete, spawn the subagents.

NEVER spawn subagents based on the first decomposition that comes to mind
without completing Steps 1-3 first.
"""

The Enumeration-First Prompt Pattern

research_request = """
Research task: Analyze the impact of AI on creative industries.

Before creating your research plan:

1. First, explicitly list every category of creative industry you can identify.
   Do not skip any. Consider: what types of work do humans do creatively?
   
2. For each category you list, mark whether your plan will cover it.
   Every category must be covered — if not, add coverage.

3. Then create your final research plan with subagent assignments.

Important: 'Creative industries' is a broad category. Your final report
should cover ALL of: visual arts, music, writing/publishing, film and video,
game design, architecture, fashion, and any others you identify.
"""

Recognizing and Escaping the Trap Mid-Run

Sometimes the trap is discovered after subtasks have already started. If the coordinator receives a synthesis result and realizes coverage is incomplete:

coordinator_review_prompt = """
You have received research findings. Before synthesizing the final report,
review the coverage:

Original task: {original_task}

Research completed:
{summary_of_completed_research}

COVERAGE AUDIT:
- List the complete scope of the original task
- For each scope item, check whether the completed research covers it
- Identify any scope items with no coverage

If you find gaps:
- Spawn additional subagents to cover the missing areas
- Do not synthesize until all scope items have coverage
"""

This mid-run check catches scope gaps before they reach the user.

Key Takeaways

Coordinator trap = narrow decomposition of broad topic
Subagents executing correctly + incomplete output = coordinator’s decomposition failure
Enumerate scope FIRST, decompose SECOND
Coverage check: every scope item must map to at least one subtask
Mid-run audits can catch gaps before synthesis
The exam always shows working subagents with incomplete aggregate output — look at the coordinator