A supplier investigation needs contract terms, payment history and company identity checks. One long prompt may bury dependencies; a large swarm of workers can duplicate research and return incompatible evidence. Before splitting work, define the decision the investigation must support and identify which facts can be collected independently. A useful decomposition for Anthropic CCA-F separates parallel evidence gathering from dependent judgments, specifies a verifiable result for each task and makes the coordinator responsible for unresolved gaps. The number of subagents is an implementation detail, not a measure of architectural quality.
Identify the final decision before splitting work
A supplier review may end with an approval, conditional approval or escalation. Before creating subtasks, establish what evidence the decision requires: company identity, contract terms, payment history, sanctions status and relevant policy. This gives the workflow a definition of completeness. Without it, workers can each produce a convincing report that leaves a critical decision input uncovered. Decomposition starts from the output contract and works backward to the evidence, rather than beginning with whatever list of tools or specialist agents happens to be available.
The contract should distinguish facts that can be gathered independently from judgments that require integration. Retrieving financial statements and checking sanctions databases may run separately. Deciding whether an unresolved entity-name mismatch is acceptable depends on comparing those results. A useful plan states which outputs are needed for the final decision and how uncertainty will be represented. That reduces the chance of a coordinator accepting a polished but incomplete synthesis simply because every worker claims its assigned portion is finished.
Use fixed decomposition for stable procedures
Some processes have predictable stages. An invoice review might always require identity checks, line-item extraction, policy validation and human approval above a threshold. A fixed decomposition is useful here because the stages are stable, testable and subject to governance. The application can define their prerequisites and enforce execution order without relying on the model to invent a new plan for each receipt. Workers can still apply judgment within their assigned stage.
Fixed workflows also simplify measurement. The team can compare error rates at extraction, validation and approval and see where failures originate. However, rigid steps become inefficient when tasks differ dramatically. A trivial receipt might require less investigation than a complex multi-currency contract. The architect should decide which controls are mandatory across every case and where optional analysis can be conditionally invoked without altering the required acceptance gates. Predictability is valuable, but not at the cost of irrelevant work.
Use dynamic decomposition for uncertain research
Exploratory investigations often cannot be fully specified in advance. A researcher examining an unfamiliar software incident might discover that the principal cause is in network configuration rather than the application code. Dynamic decomposition allows the coordinator to create a targeted network task after reviewing early evidence. This can be more efficient than always dispatching separate code, network, storage and security workers. The coordinator must still set limits on what new tasks can be created and which tools they may use.
Every dynamically created task should have a clear objective, relevant evidence, a result contract and a stopping rule. If the coordinator invents a new worker merely because a previous worker disagreed, the process can spiral into expensive repeated analysis. Track task identity, parent dependency and the question that remains unresolved. Dynamic planning is strongest when new subtasks respond to a specific information gap, not when the model is free to proliferate agents without demonstrating what those agents will add.
Make dependencies visible
Imagine a supplier review in which the payment audit is meaningful only after the team identifies the correct legal entity. Running payment analysis before resolving that identifier risks drawing conclusions from another company with a similar name. Represent this dependency explicitly: identity confirmation precedes financial analysis; independent regulatory research may proceed in parallel. The application should not let a subagent’s fluent narrative erase the missing prerequisite.
A dependency graph need not be complicated software. Even a small manifest listing prerequisite evidence, worker, accepted output and next consumers can prevent many mistakes. For complex systems, durable task states make retries and handoffs safer. If the identity stage fails, dependent workers should receive a blocked or incomplete status rather than fabricated identifiers. The coordinator can explain what evidence is missing and decide whether manual review is required before further steps are authorized.
Pass enough context, but not everything
A specialized worker needs a concise task description and the source references required for its assignment. Giving every worker the full conversation and every retrieved document increases context cost and the risk of unrelated information affecting its result. A sanctions reviewer may need the verified entity name and jurisdiction, not private payment details. A payment auditor needs authorized ledger records, not an unrestricted copy of public research comments. Context isolation should follow the task boundary and access policy.
Too little context creates a different failure. If the contract analyst never receives the effective date, it may interpret an expired clause as current. Define essential inputs for each handoff and validate their presence before execution. Record source versions and unresolved assumptions with the task. This complements subagent orchestration: workers should transmit bounded evidence and limitations, while the coordinator retains responsibility for decisions spanning those boundaries.
Design worker outputs for aggregation
A worker should return observations, supporting source IDs, confidence qualifiers, outstanding questions and a status. “Approved supplier” is often too coarse; it hides whether the worker evaluated every required condition or found a single encouraging fact. For example, an identity worker can report the legal name, registration number, source documents and mismatch status. The coordinator can then compare that record with financial and contract findings without interpreting loose prose as a complete decision.
Aggregation requires recognizing inconsistent definitions. One worker may report annual revenue in calendar years while another uses fiscal-year results; one may quote a currency at transaction date and another at report date. A structured schema allows the coordinator to detect incompatibilities instead of averaging them. Do not force workers to agree through repeated prompts when they rely on different valid definitions. The correct synthesis may preserve both measurements and explain the basis for comparison.
Limit task proliferation and duplicated research
More workers can increase coverage but also multiply calls, contradictory outputs and repeated retrieval. Use a bounded task budget and stop creating subtasks when further searches no longer address a meaningful question. A coordinator that launches five versions of the same document search does not gain independent evidence if each worker reads the same press release. Track which sources have already been reviewed and which unresolved requirements remain.
Evaluate dynamic decomposition against a simple baseline. Compare the completeness of the final decision, elapsed time, resource consumption and human corrections. If a fixed two-stage workflow handles ordinary cases reliably, reserve dynamic investigation for exceptions. The aim is proportional complexity: deploy specialization only when the problem contains separable work or uncertainty that benefits from independent exploration. The number of subagents is an architectural cost, not a badge of sophistication.
Handle blocked tasks without pretending completion
For each blocked branch, the coordinator needs the worker failure envelope and completed-source coverage, not merely a blank findings list. When the larger case is paused, durable session checkpoints retain which tasks were accepted and which still require evidence. This lets unaffected branches finish without permitting dependent decisions to treat an unavailable prerequisite as successful work.
A worker may be unable to access a document or may time out after collecting partial evidence. Its status must propagate to the coordinator along with what it actually verified. Do not substitute a confident answer from another agent unless that answer addresses the same requirement with authorized, independent evidence. If a required identity source is unavailable, the correct result may be a hold for manual verification rather than automatic approval.
Retries should follow the failure type. A transient read can often be repeated; an attempted privileged action with ambiguous outcome needs status reconciliation before another call. Persist task identifiers and evidence snapshots where appropriate so interrupted jobs can resume without losing completed work. This matters in long-running research pipelines where a failure near the end should not trigger a fresh unbounded exploration of every source. Good decomposition makes partial progress visible and recoverable.
A supplier investigation with controlled expansion
A company receives a supplier application. The coordinator requires identity, policy eligibility, contract terms and financial risk evidence. Identity and public policy research begin independently. Identity finds a mismatch between the trading name and registration record, so the coordinator creates a narrow follow-up task to resolve the legal entity before financial analysis proceeds. The contract worker identifies a clause whose effective date requires confirmation. Instead of generating an approval prematurely, the system records both evidence gaps and routes them to the responsible reviewer.
Once identity is confirmed, the payment auditor runs against the correct entity and returns an evidence-backed finding. The coordinator integrates those findings, preserves caveats and makes a recommendation within the organization’s approval process. This is effective decomposition because tasks reflect real dependencies and distinct sources of information. Fixed gates protect mandatory checks; dynamic tasks pursue discovered questions. Neither mechanism gives a model authority to bypass unresolved evidence merely to produce a neat final answer.