Partition research without shrinking the question

Learning objective

Partition broad investigations into complementary, sufficiently complete subtasks without narrowing the original question.

Core idea

For an architect coordinating broad research, decomposition succeeds when the combined assignments can answer the original question. Several detailed reports can still fail if they all examine one convenient part of it. The coordinator must preserve the user’s decision criteria before distributing work.

Anthropic’s research-system account describes delegation failures caused by vague assignments: workers duplicated searches, investigated different time periods, and left gaps. Its remedy is to give each worker an objective, expected output, guidance on tools and sources, and clear boundaries. These instructions make specialization useful without making the overall investigation narrower.

How it works

Extract obligations before choosing workers

Start with the question the user actually needs answered. Identify the decision, alternatives under consideration, comparison criteria, scope qualifiers, and required evidence. Distinguish an explicit exclusion from something the user never specified. If a time period or geography materially changes the answer, preserve it across assignments or ask for clarification rather than letting workers choose incompatible assumptions.

Build an application-owned coverage matrix that maps each requirement to an investigation and an expected evidence type. This matrix is a planning artifact, not a named Agent SDK option. It makes omissions visible before any calls begin. Include relationships among dimensions: “which approach should we choose?” often needs a cross-cutting trade-off analysis, not just independent descriptions.

For a support-agent deployment comparison, the matrix might look like this:

These are example application roles, not built-in Anthropic agents. The coordinator retains responsibility for the recommendation. If no assignment covers a requested dimension, the plan is incomplete even when every assigned worker finishes successfully.

Choose a complementary partition

Choose the partition from how the user’s criteria divide:

Keep the partition understandable enough that the coordinator can explain why each worker exists. Documentation, repository evidence, and operational reports can all describe the same dimension, so three source collections are not automatically three complementary answers.

Anthropic distinguishes sectioning from voting. Sectioning divides independent subtasks; voting repeats the same task to obtain multiple perspectives. Deliberate overlap can test a critical disputed claim, but accidental duplicate searches consume resources without improving coverage. Assign duplication a specific purpose and distinguish it from unfilled requirements.

Do not parallelize a dependency as though it were independent. If resource estimates depend on a workload definition, establish that definition first or send the same explicit assumptions to every affected worker. When one investigation discovers an assumption is invalid, route that update through the coordinator rather than allowing inconsistent estimates to accumulate.

Write an assignment that preserves the whole

In the Agent SDK, use AgentDefinition.prompt for stable specialist instructions and the Agent tool’s prompt string for this request’s scope. Include the relevant original requirements directly rather than referring to an earlier discussion; see 1.2.1 for the context and routing boundary.

A worker assignment should answer five questions:

• What part of the original decision does this investigation serve?
• Which alternatives, dates, and assumptions must remain consistent?
• Which sources and tools should supply evidence?
• What findings, supporting locations, and limitations must the report return?
• Which neighboring questions belong to other workers, and how should dependencies be reported?

Give the worker enough breadth to discover material unknowns. “Find reasons approach A is best” biases an open comparison before research. “Evaluate approach A against the shared criteria, including disqualifying constraints” preserves the decision. An exclusion that prevents duplicate work should still permit a worker to flag evidence that affects another criterion.

Start broad within the assigned scope

Anthropic recommends starting research with short, broad searches, evaluating the landscape, and narrowing afterward. Apply that principle inside each assignment. A specialist can first identify the relevant mechanisms and evidence categories, then pursue the decision-critical details. Narrow initial searches can miss the terminology or alternative the coordinator should have considered.

Do not mistake broad discovery for an unlimited assignment. Keep the user’s decision fixed while allowing the research path to adapt. If the landscape reveals an omitted alternative, the worker returns that finding to the coordinator. The coordinator then decides whether to revise the plan and add an investigation.

At result intake, compare the reports against the original matrix. Mark obligations supported, unresolved, or unsupported by the supplied evidence. Task completion and evidence completeness are separate checks. See 1.2.4 for repairing the remaining gaps through targeted re-delegation and synthesis.

Key distinctions and trade-offs

Complementary versus merely different: Different agent names or source lists do not guarantee that assignments cover different obligations. Judge the union of their scopes against the original request.

Bounded versus biased: A bounded task limits an investigation’s responsibility. A biased task removes alternatives or unfavorable evidence that the user’s decision requires.

Independent versus sequential: Parallel work helps when investigations can proceed with shared assumptions. A dependency on an unknown earlier result calls for sequencing or explicit provisional assumptions.

Finer partitions give workers focused contexts, but increase handoffs and integration work. Coarser partitions reduce coordination overhead, but can leave one worker juggling too many criteria. Choose a granularity that preserves evidence and interfaces between investigations rather than maximizing agent count. The count follows the partition, not the other way round; see 1.2.3 for scaling effort to the request.

Architecture / decision scenario

A multi-agent research application receives: “Compare two ways to deploy our customer-support agent, including resilience, data handling, and resource implications.” The coordinator initially proposes three workers, each investigating a different failure mode. That plan appears parallel but drops data handling and resource implications entirely.

The coordinator replaces it with criterion-based assignments using a shared workload definition and the same two alternatives. The operations researcher investigates failures and recovery, the data researcher checks the stated information-handling requirements, and the cost researcher evaluates comparable resource assumptions. Each assignment requires evidence locations, uncertainty, and dependencies.

The data researcher discovers a deployment constraint that changes the feasible recovery options. It returns the finding to the coordinator, which forwards it to operations and requests a focused update. The coordinator permits a second check of that constraint because the recommendation depends on it, while avoiding duplicate general searches. Before synthesis, it checks that every requested criterion has evidence for both alternatives. The plan remains bounded without silently converting a deployment decision into a failure-mode survey.

Common exam traps

• Dividing one easy dimension among many workers and calling that comprehensive research.
• Answering a coverage gap with more subagents. Extra workers inside a narrow decomposition re-examine the same slice; the dropped criteria stay unassigned until the coordinator repairs the partition.
• Assigning different source types without mapping them back to required criteria.
• Giving workers incompatible periods, alternatives, or workload assumptions.
• Removing a relevant alternative while making an assignment more specific.
• Treating accidental duplicate searches as independent validation.
• Equating completed assignments with a sufficiently complete answer.

Where the exam guide differs

What to remember

Preserve the decision before dividing the work. Map every requirement to an owner and evidence expectation, share essential assumptions explicitly, and let specialists flag cross-cutting discoveries. Complementary subtasks reduce duplication while their combined scope still answers the whole question.

Knowledge check

References

• Anthropic, How we built our multi-agent research system
• Anthropic, Building effective agents
• Anthropic, Subagents in the SDK
• Anthropic, Create custom subagents
• Anthropic, Claude Certified Architect – Foundations Exam Guide