Delegate outcomes with testable evidence criteria

Learning objective

Define delegated goals, evidence expectations, and quality criteria rather than prescribing every procedural step.

Core idea

For an architect delegating research, a good assignment defines what the specialist must establish and how the coordinator can judge the result. Specify the question, the evidence needed to answer it, and the criteria for a sufficient report. Let the specialist choose searches and revise its approach when evidence changes the next useful action.

This is a concrete prompting pattern. In the Agent SDK, AgentDefinition.prompt carries the specialist’s reusable role and behavior; the Agent tool’s prompt carries the particular assignment. description helps Claude decide when to invoke the specialist. See 1.3.1 for the full contract. This lesson focuses on what those instructions demand, rather than repeating their configuration.

How it works

State a goal the report can answer

Anthropic’s multi-agent research account says each subagent needs an objective, an output format, guidance on tools and sources, and clear task boundaries. Its examples show that vague delegation causes duplicated work and missed scope. Replacing procedural instructions with a goal therefore does not mean reducing an assignment to “research this topic.”

A useful goal identifies the decision and the specialist’s contribution. For example, ask a deployment researcher to establish which documented deployment modes satisfy a supplied isolation requirement. The report must let the coordinator choose between the relevant modes. Asking the specialist to “read three pages and summarize them” defines activity without establishing whether the decision is answered.

Put reusable decision responsibilities in AgentDefinition.prompt, then include the specific requirement and candidate modes in the Agent call’s prompt. A fresh subagent does not receive the parent’s conversation history, so the assignment must carry the required context; see 1.3.3. Goal-based delegation still depends on complete inputs.

Specify what counts as evidence

Give the specialist a source standard and a rule for unsupported claims. For a platform comparison, require primary product documentation for support claims, a precise supporting passage for each deciding constraint, and an explicit unresolved status when the source does not settle the question. These are example assignment criteria, not built-in SDK fields.

The mechanism is instruction text in AgentDefinition.prompt and the Agent tool’s prompt. There is no automatic evidence check implied by choosing a research-oriented description. The parent receives the subagent’s final report, while intermediate tool calls and results stay in the subagent’s conversation. Require the final report to include the evidence the next stage needs.

Separate the report’s finding from its attribution metadata; see 1.3.4 for the handoff structure. Here the deciding question is whether the evidence actually establishes the finding. A source URL alone does not tell the synthesizer which condition the source supports. A plausible conclusion without supporting content does not meet an evidence expectation merely because the report has a bibliography.

Make quality criteria observable

Anthropic describes evaluating research with factual accuracy, citation accuracy, completeness, source quality, and tool efficiency. Translate those dimensions into checks on the delegated result. For example, completeness means that the report addresses every supplied deployment mode and the isolation requirement. Citation accuracy means that the referenced passage supports the specific compatibility claim attached to it.

A criterion such as “be thorough” does not define a stopping condition. A clearer assignment requires a decision for each candidate, its supporting evidence, material contradictions, and unresolved questions. That gives the coordinator a basis for accepting the report or requesting a targeted follow-up. See 1.2.4 for the refinement loop; the criterion is the input to that loop, not another fixed research stage.

Specify the output format alongside the criteria. The SDK documents specialist prompts that request concise findings and specific improvements. In your research assignment, a compact comparison organized by candidate makes omissions visible. The format is useful because it exposes coverage, not because a table guarantees a correct answer.

Leave the investigative route adaptable

Anthropic’s research prompting strategy favors useful heuristics over rigid rules: evaluate sources, adjust searches when information changes, and choose depth or breadth according to the question. A specialist may discover that a candidate has been renamed or that a support statement applies only to one deployment mode. The assignment should let it pursue that discovery instead of forcing the next unrelated step in a script.

Source guidance and task boundaries still belong in the assignment. “Use primary product documentation; compare the supplied modes; report unsupported cases” gives the specialist discretion over search queries while preserving the decision. “Search these exact phrases in this exact order and stop after five results” can produce a completed sequence with an unanswered question.

Keep permission and tool control separate from this discretion. AgentDefinition.tools restricts the specialist’s available tools. For repository-based analysis, the SDK documents Read, Grep, and Glob as a read-only combination. A goal-based prompt cannot grant an unavailable tool, and flexibility does not justify widening access. A researcher can adapt its approach within its assigned tools and scope.

Key distinctions and trade-offs

The assignment’s clauses serve different purposes:

A precise result contract leaves room for exploration while reducing ambiguous acceptance. A fixed sequence can be appropriate for a mandatory workflow, but it should not substitute for evidence criteria in an open-ended investigation. See 1.4.1 for enforced prerequisites; this lesson’s adaptable research instructions are not an authorization mechanism.

More detailed instructions also have a cost: irrelevant constraints can narrow the investigation prematurely. Keep the required coverage and source standard explicit, and avoid prescribing a search route that assumes the answer. The specialist must be able to discover that the proposal’s premise is unsupported.

Architecture / decision scenario

A multi-agent research system evaluates a developer platform for a customer with a strict isolation requirement. Its draft Agent prompt says: “Search these three phrases, read the first five results, and summarize each.” You must choose whether to retain this script or replace it with a result contract. The script can finish without establishing which deployment modes meet the requirement, so you choose the result contract.

The coordinator delegates deployment compatibility to one specialist and access-control requirements to another. Each Agent prompt contains the shared requirement, the candidate modes, and the specialist’s distinct scope.

The deployment specialist’s AgentDefinition.prompt defines the goal as determining which candidate modes meet the isolation requirement. It requires primary documentation as evidence and a candidate-by-candidate decision, with contradictions and unresolved support claims identified, as the quality criteria. The coordinator requests a report organized by candidate, finding, evidence, and limitation. It does not prescribe search phrases or a page-reading order. The access-control specialist receives the same evidence standard for its own question.

The deployment researcher finds a support statement limited to one mode. It changes its search approach to investigate that limitation rather than completing an arbitrary sequence. Its report marks another candidate unresolved because the available documentation does not settle the isolation claim. This meets the explicit unresolved-status criterion even though it does not produce an unconditional recommendation.

Before synthesis, the coordinator checks candidate coverage and claim-to-source support. A missing candidate triggers a targeted assignment; an unresolved claim stays unresolved unless further evidence settles it. The synthesizer then compares the supported findings against the customer’s requirement. The architecture rewards a defensible answer, rather than treating completed searches or confident prose as proof of quality.

Common exam traps

• Replacing a procedure with a vague topic and calling it goal-based delegation.
• Defining success as a count of searches instead of evidence sufficient for the decision.
• Assuming description enforces the specialist’s evidence standard.
• Expecting the parent to inspect all intermediate research without requesting evidence in the report.
• Prescribing a conclusion that prevents the specialist from reporting uncertainty.
• Confusing investigative flexibility with permission to exceed tools or task boundaries.

What to remember

Use AgentDefinition.prompt and the Agent call’s prompt to define goals, evidence expectations, quality criteria, scope, and output. Let the specialist adapt its route within those constraints. Judge the returned report by what it establishes and supports, then target follow-up work at the remaining gaps.

Knowledge check

References

• Anthropic, Subagents in the SDK
• Anthropic, How we built our multi-agent research system