Skip to lesson

AnthropicClaude Certified Architect – Foundations2026-101.3 Configure subagent invocation, context passing, and spawning

Keep evidence attached to its attribution

Completed by the knowledge check

Learning objective

Preserve evidence separately from source URLs, document names, page locations, and other attribution metadata in structured handoffs.

Core idea

For an architect building research synthesis, a source URL is an address, not the evidence that supports a claim. A document title identifies a source but says nothing about which passage supports the comparison. A prose conclusion can preserve the answer while losing the content and location needed to check it.

Structure the handoff so the supporting content and its attribution occupy distinct fields linked by a stable record identifier. The receiving specialist can then compare evidence, trace an interpretation to its passage, and preserve citations through synthesis. Keeping metadata separate does not mean sending disconnected lists that rely on positional order.

The exam guide explicitly calls for structured separation of content from source URLs, document names, and page numbers. Anthropic’s research-system account describes a citation stage that processes research and documents to find specific citation locations. The schema in this lesson is an application design for that requirement, not a built-in Agent SDK handoff format.

How it works

Preserve the supporting content

Ask the research specialist to return the passages needed by the next agent, not only a verdict and bibliography. Put this expectation in its AgentDefinition.prompt and include request-specific output requirements in the Agent tool’s prompt. See 1.3.1 for role contracts and 1.3.3 for complete task inputs.

Anthropic’s hallucination guidance recommends direct quotes for grounding document analysis and citations that make claims auditable. It also recommends checking a generated claim against a supporting quote and retracting it when support cannot be found. A handoff that discards the quote makes that check dependent on reopening the source.

Preserve the difference between a passage and an interpretation. If a policy says deletion happens after account closure, a researcher might infer a retention duration under particular conditions. Put the observed passage in an evidence field and the derived conclusion in a finding field. The synthesis specialist needs enough content to check whether the interpretation applies to the question, including any stated exception or qualification.

Keep attribution in explicit fields

For supplied-document analysis through the Messages API, use its generally available Citations feature: set "citations": {"enabled": true} on each document block, enabling all or none in a request. All active models support citations. Returned citation records separate cited_text from document_index, document_title, and location. PDFs use type: "page_location" with 1-indexed start_page_number and exclusive end_page_number; plain text uses type: "char_location" with 0-indexed start_char_index and exclusive end_char_index.

Map cited_text into handoff content and the document identity and location into attribution. The API extracts the cited text and guarantees valid pointers into supplied documents; that does not prove the resulting claim follows from the passage. cited_text does not count toward output tokens. These are Messages API citations, not blocks emitted by the Agent SDK’s Agent tool.

Citations are incompatible with structured outputs: enabling citations alongside output_config.format returns a 400 error. Build the JSON handoff from returned citation records in your application instead of requesting that combination. PDF citations require extractable text; image-only scans are not citable.

A source record should identify the document and the location of the supporting content. For a PDF, that might be the supplied document name, URL, page, and section. For a web page without page numbers, use the section or another meaningful locator instead of inventing a page. If no URL exists for a supplied internal file, retain its file identity and mark the missing URL explicitly.

The following JSON illustrates an application-defined handoff using a hypothetical policy excerpt:

{
  "evidence": [
    {
      "id": "E1",
      "content": {
        "quote": "Case records are deleted 30 days after case closure."
      },
      "attribution": {
        "url": null,
        "document": "Supplied support retention policy",
        "page": 4,
        "section": "Case records"
      }
    }
  ],
  "findings": [
    {
      "claim": "The policy states a 30-day period after case closure.",
      "evidence_ids": ["E1"]
    }
  ],
  "gaps": ["No supplied passage specifies backup retention."]
}

The field names evidence, content, attribution, evidence_ids, and gaps are your application’s contract. The example’s duration and page number are invented inputs, not Anthropic product facts. The null URL records an unavailable source address rather than fabricating a public link.

The E1 identifier connects the finding to its evidence regardless of array order. The quote preserves content; the attribution identifies where it came from. Neither the existence of E1 nor valid JSON establishes that the passage supports the claim. The receiving agent still has to evaluate that relationship.

Pass the complete records to the synthesis specialist through its task prompt. Require each material conclusion to identify supporting evidence records and preserve the corresponding attribution. The SDK guide states that the parent receives the specialist’s final report but can summarize it in its response. If the user-facing answer must preserve that report verbatim, the guide recommends an explicit instruction in the main query() prompt or systemPrompt option.

Verbatim preservation and evidence validation solve different problems. A verbatim result can still contain an unsupported claim. A concise paraphrase can remain accurate if it retains the relevant qualifiers and evidence links. Decide which deliverable you need and specify it in the coordinator’s instructions.

Anthropic’s research architecture also uses a citation stage that locates support in the research documents. When your design has such a stage, give it both the report and the underlying evidence. A citation agent cannot validate a deletion claim using only a list of policy titles. The title names a candidate source, while the passage supplies the statement it must check.

For large results, Anthropic describes artifacts that persist outside the lead agent’s conversation and lightweight references returned to the coordinator. Store the complete evidence records in the artifact rather than storing only a summary. Pass the necessary excerpts directly when the next agent cannot retrieve the artifact. See 1.3.3 for that retrieval boundary.

Key distinctions and trade-offs

Evidence versus metadata: Evidence contains the statement or observation you reason from. Metadata identifies its source and location. Both are necessary for an attributable claim, but neither replaces the other.

Structured representation versus verified support: JSON makes the relationships explicit and inspectable. It does not prove authenticity, correct attribution, or entailment. Check that a quote belongs to the named document and supports the specific claim made from it.

Missing evidence versus negative evidence: A reviewer that found no backup clause has established a gap in the supplied material. It has not established that backups are deleted on the same schedule or that backups do not exist. Preserve that gap as a separate output rather than converting it into a confident comparison.

More detailed excerpts increase handoff size but reduce dependence on retrieval and protect qualifications. Shorter summaries reduce size but can erase the exact condition that changes the decision. Preserve the minimum sufficient passage, its relevant qualifiers, and its attribution together.

Architecture / decision scenario

A multi-agent research system compares support-record retention in two supplied vendor policies. The Vendor A reviewer returns “Vendor A: 30 days” with only a bare source URL as evidence. The Vendor B reviewer returns an excerpt that applies only after case closure, with a document name and page. The coordinator needs a comparison for active cases and backups, so the first result lacks supporting content and the second leaves scope gaps.

You require each reviewer to return evidence records with content separate from attribution, findings linked through evidence_ids, and explicit gaps. The first reviewer retrieves the supporting passage and its section. The second preserves the closure condition in the quote rather than shortening it to the number. Both retain the supplied document names and page locations where available.

The synthesis specialist receives those records directly and compares the stated scope of each clause. It can report the documented periods for closed cases while marking active-case and backup treatment unresolved. The citation stage checks the quoted clauses against their recorded document locations and keeps the evidence links attached to the final comparison.

If the coordinator instead strips all records down to “A is better” plus two URLs, the attribution machinery loses the deciding condition. Repair the structured handoff before accepting synthesis; see 1.2.4 for the broader refinement loop. Accurate comparison requires preserving what was said, where it was said, and what the material leaves unanswered.

Common exam traps

  • Passing source URLs as if they were already-read evidence.
  • Mixing document titles, page numbers, and claims into unlabelled prose.
  • Separating content and metadata into arrays without stable associations.
  • Treating parseable JSON as proof of a correct claim.
  • Removing conditions from quotes while preserving only a deciding number.
  • Converting an unlocated clause into a confident negative conclusion.

What to remember

Preserve supporting content, source identity, and precise location as distinct but connected fields. Keep observations separate from interpretations and gaps. Carry those relationships into synthesis and citation checks; a bibliography alone cannot reconstruct evidence the handoff discarded.

Knowledge check

References