Bad Task Decomposition

Issue: Agent splits the task into wrong subtasks, causing missed work.

Frequency: Occasional

Symptoms

• Subtask list does not cover acceptance criteria.
• Subtasks overlap or duplicate effort while a required step (e.g., compliance sign-off) never appears as its own line item.
• Agent marks the parent task complete after finishing only the subtasks it generated, even though the original request listed additional deliverables.
• A single subtask silently bundles two distinct systems or owners (e.g., “update ERP and notify finance”), so only one half actually executes.
• Plan granularity is inconsistent: trivial steps get their own subtask while a critical multi-part step is folded into one line.

Root Cause Task requests typically arrive as flowing prose rather than an itemized list, which makes it easy for decomposition to silently fold two distinct requirements — like document verification and payment-terms setup — into a single subtask that only actually performs one of them. Because no acceptance-criteria extraction step runs before decomposition, there is no structured checklist to validate the subtask list against, and the decomposition prompt tends to gravitate toward whichever tools and APIs are readily available rather than the full scope of the request. With only a single decomposition pass and no critic reviewing coverage before execution begins, and no template forcing one subtask per system when a task spans multiple owners, gaps in the plan are indistinguishable from a complete one until missing work surfaces much later.

Example

A procurement automation agent is asked to "onboard new vendor Acme Fasteners: create the vendor record, verify the W-9 and insurance certificate, set payment terms, and notify the requesting department." The agent decomposes this into two subtasks — "create vendor record in ERP" and "notify requesting department" — because its decomposition prompt folded document verification and payment-terms configuration into an implicit part of "create vendor record" that the ERP-creation tool call never actually performs. The vendor record gets created with default net-30 terms and no compliance documents on file, the notification goes out, and the missing insurance certificate isn't discovered until a vendor-risk audit three months later.

Contributing Factors

• Task spec is written as flowing prose rather than an itemized list, making it easy for decomposition to fold multiple requirements into one subtask.
• No acceptance-criteria extraction step runs before decomposition, so there is nothing to check the subtask list against.
• Decomposition prompt biases toward the tools/APIs the agent has readily available rather than the full scope of the request.
• Single-pass decomposition with no critic or review step before execution begins.
• Task spans multiple systems or owners (ERP, compliance, accounts payable) and no template forces one subtask per system.

Eval Recipes

Test Cases

Metrics

Mitigation Strategies

Prevention

1. Acceptance-Criteria-to-Subtask Traceability Matrix: Before execution, require every acceptance criterion extracted from the task spec to map to at least one subtask in the decomposition. Generate this mapping as a structured table (criterion_id → subtask_ids) and block plan approval if any criterion has zero mapped subtasks.
2. Decomposition Critique Pass: Run a separate critic model/prompt that reviews the proposed subtask list against the original task spec, checking for coverage gaps, overlapping/redundant subtasks, and ambiguous scope boundaries. The critic must explicitly answer “does this decomposition fully cover the spec?” before the plan proceeds to execution.
3. Template-Based Decomposition for Known Task Types: For recurring task classes (e.g., “migrate a table,” “onboard a customer”), maintain vetted decomposition templates listing the standard required subtasks. The agent’s generated plan is diffed against the template, and missing template steps must be justified or added.

Detection & Response

1. Coverage Gap Scanner: At the point the agent proposes a final answer, automatically re-run the criteria-to-subtask matrix check against actual completed work (not just the plan). Any criterion without a corresponding completed subtask blocks finalization and forces a replan cycle.
2. Post-Hoc Decomposition Audit Sampling: Sample a percentage of completed tasks weekly and have a reviewer (human or LLM-judge) score decomposition completeness against the original spec, producing a completeness_score used to track drift over time.
3. Downstream Rework Rate Tracking: Track how often completed tasks require a follow-up correction because a subtask was missed. Rising rework rate for a given task type is a leading indicator of decomposition quality degradation and should trigger a review of the planner prompt or template.

Architecture Patterns

1. Planner-Critic Loop: Two-stage pipeline where a plan generator produces a structured subtask list {subtask, maps_to_criterion, dependencies} and a separate critic stage validates it against the parsed spec before the plan is unlocked for execution.
2. Spec-Grounded Decomposition Service: Parse the task spec into a structured criteria list (via schema extraction) first; the decomposition engine consumes this structured list rather than raw free text, so coverage checking is mechanical rather than inferred.
3. Decomposition Versioning & Diffing: Persist every subtask list with a version number; when a plan is revised mid-task, diff against the prior version to surface any acceptance-criteria-mapped subtask that was silently dropped.

Metrics

1. criteria_coverage_rate_percent: Target: 100%; Alert threshold: < 95%
2. decomposition_rework_rate_percent: Target: < 5%; Alert threshold: > 10%
3. subtask_missing_from_plan_count: Target: 0 per task; Alert threshold: > 0
4. critic_gap_catch_rate_percent: Target: > 90% of injected gaps caught in eval; Alert threshold: < 75%

Alerts

1. Uncovered Acceptance Criterion (P1 - Critical): Condition - plan approved for execution with an acceptance criterion unmapped to any subtask. Action: Block execution, force replan, notify task owner.
2. Elevated Rework Rate (P2 - Warning): Condition - rolling 7-day rework rate for a task type exceeds 10%. Action: Review decomposition template/critic prompt for that task type.
3. Critic Disagreement Spike (P3 - Info): Condition - critic flags an unusually high fraction of plans as incomplete in a given period. Action: Investigate planner model/prompt drift before tightening thresholds further.

Production Signals

Key Metrics

Alerts

References

• MAST
• Note: Multi-agent system failure taxonomy with system design, inter-agent misalignment, and task verification failures.