What you will be able to do
- Explain how source attribution gets lost when subagent findings are compressed into free-form summaries
- Design a subagent output contract that pairs every claim with its source and a supporting excerpt
- Handle conflicting statistics from credible sources by keeping both values with attribution and letting the coordinator reconcile them
- Require publication or collection dates in structured outputs so differences over time read as trends, not contradictions
- Structure a synthesis report into established and contested findings, and render each content type in its own format
Key concept
Claim-source mapping — A structured record that ties each individual claim to the document or URL it came from and the exact excerpt that supports it. Every step of a multi-agent pipeline passes it along and merges it, so a claim in the final report can still be traced back to the passage behind it.
1.Why attribution disappears during summarization
Provenance is the link between a claim and the passage that supports it, and a multi-agent research pipeline breaks that link by default. In an orchestrator-worker design, a lead agent sends subagents out to read sources and report back, then combines what they found. Each handoff compresses the findings. If a subagent returns a fluent paragraph such as "the market grew strongly, driven by enterprise demand", the claim makes it through and the reference behind it does not. Once the coordinator merges five of those paragraphs into one report, no step in the pipeline can say which document said what.
Anthropic's own research system shows the alternative. Subagents return findings to a LeadResearcher, which synthesizes them. After that, a separate CitationAgent processes the documents and the report together so that claims are tied back to specific locations. That design only works because the documents and the claims are both still available when citation happens. A summary that has already thrown away its sources leaves nothing for a later step to attach a citation to.
The Messages API citations feature shows what a preserved mapping looks like. Instead of one block of prose, the response is split into text blocks, and each claim carries a list of citations pointing to exact locations in the supplied documents. A plain-text citation records the quoted text, which document it came from and the character range. That record is the unit of provenance a synthesis pipeline has to keep intact. Rewriting it as prose loses it.
{
"type": "char_location",
"cited_text": "The exact text being cited", // not counted toward output tokens
"document_index": 0,
"document_title": "Document Title",
"start_char_index": 0, // 0-indexed
"end_char_index": 50 // exclusive
}Carrying these records forward costs little. The cited_text field does not count toward output tokens, and when it is passed back in later turns it does not count toward input tokens either. Token cost is therefore a weak reason to strip citations out of intermediate results. The same loss can also happen inside one long-running agent when older turns are compressed; that mechanism belongs to context compaction and case-facts blocks.
Checkpoint 1 of 8· Fill the gap
A subagent passes this citation record to the coordinator, so the claim stays tied to the exact passage behind it. Which field holds that supporting passage?
{
"type": "char_location",
" ? ": "The exact text being cited", // not counted toward output tokens
"document_index": 0,
"document_title": "Document Title",
"start_char_index": 0, // 0-indexed
"end_char_index": 50 // exclusive
}cited_text holds the word-for-word passage. Together with the document and character range, it makes up the claim-source unit that a pipeline must keep intact. Passing it along costs no output tokens, and no input tokens when sent back in later turns, so token cost is a weak reason to replace it with a prose summary.
2.The subagent output contract: claim, source, excerpt
Provenance survives synthesis only when subagents are required to return it. The lead agent decides each subagent's output format, and that format has to be a list of claims where every claim carries three things: the claim itself, the document name or URL it came from, and the excerpt that supports it. Delegation guidance from Anthropic's research system lists the output format as one of the things every subagent brief needs, alongside the objective, the tools and sources to use, and clear task boundaries.
The search_result content block is a good model for that record. source and title are required, and source accepts any stable identifier, whether a URL or an internal key such as kb://article-1234. Claude cites search results automatically when citations are enabled, and each citation carries the source and title you supplied. A subagent that reads a company knowledge base can therefore return results that keep their attribution all the way into the coordinator's context.
{
"type": "search_result",
"source": "https://example.com/article", // Required: Source URL or identifier
"title": "Article Title", // Required: Title of the result
"content": [
// Required: Array of text blocks
{
"type": "text",
"text": "The actual content of the search result..."
}
],
"citations": {
// Optional: Citation configuration
"enabled": true // Enable/disable citations for this result
}
}Each of the three fields is there to prevent a specific failure. A list of claims followed by a bibliography of everything consulted looks traceable, but nothing connects claim four to document two. An agent ID and a run timestamp record who produced a claim and when, not where the claim came from. A claim written as a short quotation, with no source name, still cannot be located. The reduce-hallucinations guidance asks for both parts together: have the model cite quotes and sources for each claim. The excerpt lets a reviewer check the claim, and the source lets them find it.
Enforce the contract with a schema rather than an instruction in the prompt. Structured outputs guarantee required fields and field types, so a subagent cannot leave out source on one claim out of forty. Schema mechanics are covered in enforcing structured output with JSON schemas. The synthesis agent then has a job of its own: merge the mappings rather than flatten them. When two subagents report the same claim from different documents, the merged claim keeps both sources and both excerpts.
Checkpoint 2 of 8· Check yourself
Reviewers of your research pipeline's final report cannot tell which document supports any particular claim. You are revising the output format the lead agent gives its subagents. Which change fixes the problem?
Only a per-claim pairing of source and excerpt maps each claim to where it came from. A bibliography doesn't say which claim came from which document. Agent IDs and timestamps record who produced a claim and when, not where it came from. A quotation with no source name can't be located.
“Make Claude's response auditable by having it cite quotes and sources for each of its claims.”Source: platform.claude.com
Checkpoint 3 of 8· Exam question
A market-research pipeline dispatches five research subagents, each of which reads several documents and returns a free-form paragraph summary to a coordinator agent. When the coordinator assembles the final report, reviewers find that several claims can no longer be traced to any specific source. Which change to the subagent output contract would best prevent this loss of provenance?
Correct answer: A — Have each subagent return a list of claims, each paired with a document name or URL and a supporting excerpt
- A. Correct. Each claim carries its own document name or URL plus an excerpt, so the link between assertion and origin survives when the coordinator merges the output of five subagents.
- B. A closing list names the documents a subagent consulted overall, but no claim is tied to a particular document. Because each subagent read several documents, a claim still cannot be traced to its specific source.
- C. An agent ID and a timestamp show which subagent produced a claim and when, not which document it came from. Each subagent read several documents, so the claims remain untraceable to a specific source.
- D. Quoted text sits close to the original wording, but without a document name or URL the coordinator cannot tell which document a quotation came from. The claims therefore still lack a traceable source.
3.Conflicting statistics from credible sources
When credible sources disagree, report both values, each with its own attribution, and mark the conflict. Every other choice puts a claim in the report that the evidence does not support. An average looks even-handed, but $440 million appears in no document, so it fails the basic grounding check of finding a direct quote that supports each claim. The citations feature behaves the same way by design: it will not point to documents or locations that were not supplied as sources. A figure invented in synthesis has nothing to cite.
Choosing one figure fails in a quieter way. It drops a credible data point and presents an unsettled question as settled, which works against the practice of having Claude verify information across multiple sources. Giving the model explicit permission to admit uncertainty reduces false information. In a report, admitting uncertainty means writing "$420M (press release) vs $460M (analyst report), conflicting" and not a single confident number.
Reconciliation is the coordinator's decision, and it happens before synthesis. A subagent finishing its document analysis should include every conflicting value it found, annotated with its source, and leave the choice to someone else. The coordinator has the wider view. In Anthropic's research system, the lead agent reviews returned findings and decides whether more research is needed, and it can spawn more subagents or change its strategy. Faced with an annotated conflict, it might send a subagent to find out whether the two figures measure different things, such as different fiscal years or definitions of revenue. If they really disagree, it passes both values on to synthesis.
Checkpoint 4 of 8· Check yourself
Halfway through a batch of filings, a subagent finds that two credible documents give different headcounts for the same company. In this pipeline, what should the subagent do?
The subagent should finish its work and include every conflicting value with its source attached. Reconciliation belongs to the coordinator before synthesis, because the coordinator has the wider view. It can review the findings and send out more subagents, for example to check whether the two figures measure different things.
“The LeadResearcher synthesizes these results and decides whether more research is needed—if so, it can create additional subagents or refine its strategy.”Source: www.anthropic.com
Checkpoint 5 of 8· Exam question
During a competitive-intelligence synthesis, two subagents report different figures for a competitor's annual revenue: one cites a press release stating $420 million, the other cites an analyst report stating $460 million. Both sources are credible. How should the synthesis agent handle this discrepancy in the final report?
Correct answer: A — Report $420 million (press release) and $460 million (analyst report) for revenue, flagged as conflicting
- A. Correct. Two credible sources disagree, so the report should give each value with its source and mark the disagreement. That keeps both data points available and lets the reader judge which figure suits their purpose.
- B. Choosing the analyst figure and dropping the press release value discards a credible data point and leaves the reader unaware that another source gave a different number. Nothing in the scenario makes the analyst figure more reliable than the other source.
- C. Choosing the press release figure and dropping the analyst value discards a credible data point and leaves the reader unaware that another source gave a different number. A company's own statement is not shown to be more accurate than an independent analyst's.
- D. The midpoint is a number neither source reported, and citing both as its basis makes it look like a sourced figure. The real disagreement between two credible sources is hidden instead of shown.
4.Publication and collection dates as required fields
A difference between two figures is only a contradiction if they describe the same moment. Regional unemployment of 5.1% in one report and 6.3% in another published fourteen months later may be two accurate points on a trend. Without dates, the coordinator cannot tell those apart, so it spends effort reconciling a disagreement that does not exist, or it averages two values that should stay separate.
Anthropic's research system ran into the same topic split across time. One subagent researched the 2021 automotive chip crisis while others looked at 2025 supply chains. Figures from those two runs describe different periods of the same subject, and a synthesis agent that cannot see the period will compare them as if they described the same year.
The fix belongs in the subagent's structured output. Every extracted figure carries the publication date of its source or the date its data was collected, as a required field next to the claim, source and excerpt. Document metadata has standard places to go. When documents are passed with citations, the context field holds metadata as text or stringified JSON, and in prompts each document can be wrapped with a <source> tag and other metadata subtags. Making the date field required in the schema means the coordinator never receives a figure without one.
It removes the older data point when, with its date attached, that point is evidence of a trend. It also makes the subagent decide what to keep, which is a judgement the coordinator should make with every value in front of it. Averaging into a twelve-month figure or rounding into a range such as 5–6% fails for the same reason as averaging two revenue figures: the result is a value no source reported.
Checkpoint 6 of 8· Check yourself
Your coordinator keeps flagging 'contradictions' between subagents. For example, one reports regional unemployment at 5.1% and another at 6.3%, and it spends extra research rounds trying to reconcile them. On review, the two reports were published fourteen months apart. Which change to the subagent output schema prevents this?
A difference is only a contradiction if both figures describe the same moment. A required date field on each figure, such as document metadata carried in the context field, lets the coordinator see two points on a trend. A run timestamp only records when the agent ran. Keeping only the newest figure throws away evidence and takes the decision away from the coordinator. A merged range is a value no source reported.
“title is limited in length, so the context field is useful for storing document metadata as text or stringified JSON.”Source: platform.claude.com
Checkpoint 7 of 8· Exam question
A synthesis agent combines a subagent finding that unemployment in a region is '5.1%' with another subagent finding that it is '6.3%'. The coordinator initially treats this as a contradiction requiring reconciliation, but on closer inspection the two figures come from reports published fourteen months apart. What structural change to subagent output would prevent this kind of false contradiction going forward?
Correct answer: A — Have each subagent record the report's release date with each figure it extracts before passing it to the coordinator
- A. Correct. A release date next to each figure shows the coordinator that 5.1% and 6.3% describe different points in time, so the gap is a change over fourteen months and not a conflict between sources that needs reconciling.
- B. Replacing a reported figure with a twelve-month average blends the two readings into smoothed numbers and discards the point-in-time values. A single report often lacks that history, and the output still carries no date to explain why the sources differ.
- C. Each subagent holds only its own report, so keeping the newest figure within one subagent does nothing about the difference between two subagents' findings. The surviving figures still carry no date, so the coordinator still cannot tell a time gap from a real conflict.
- D. Rounding into ranges only gives up precision. The cause of the false contradiction is that the coordinator cannot see when each figure was published, and ranges add no timing information. They would also blur real changes between periods.
5.Established vs contested sections and fit-for-purpose formats
A synthesis report should keep what the evidence settles apart from what it does not. Findings supported consistently across sources go in an established section. Findings where credible sources disagree, or rest on a single source, go in a contested section, along with the annotated conflicts the coordinator passed through. A reader can then rely on the first section and treat the second with caution, which a single blended narrative does not allow.
Within both sections, keep each source's own characterization rather than rewording it. If a study describes its result as preliminary, or bases it on a small sample, that wording and that methodological context are part of the finding. A synthesis agent that rewrites "early survey data suggests" as "data shows" has made the claim stronger than its source does. Grounding the work in quotes protects against this: having the model extract word-for-word quotes before doing its task ties the output to what the text actually says.
Render each content type in the format that suits it, instead of converting everything to one style. Financial data belongs in tables, where figures, periods and sources sit in columns that can be compared. News developments read best as prose, because sequence and cause matter. Technical findings work best as structured lists that can be scanned item by item. The citations API makes the same distinction for input documents: custom content documents exist for content whose structure needs to be kept, rather than chunked into sentences like prose.
| Type | Best for | Chunking | Citation format |
|---|---|---|---|
| Plain text | Simple text documents, prose | Sentence | Character indices (0-indexed) |
| PDF files with text content | Sentence | Page numbers (1-indexed) | |
| Custom content | Lists, transcripts, special formatting, more granular citations | No additional chunking | Block indices (0-indexed) |
Checkpoint 8 of 8· Check yourself
A stakeholder reading your pipeline's final report says it presents everything with the same certainty. One source called its result "early survey data suggests", but the report says "data shows". A claim backed by a single study sits next to findings that five sources agree on. How should you change the synthesis step?
Keeping established and contested findings in separate sections tells the reader which claims they can rely on. Grounding the synthesis in word-for-word quotes keeps qualifiers like "preliminary" or "small sample" that rewording would remove. A uniform confident voice makes claims stronger than their sources do. Dropping contested findings hides a disagreement the reader needs to see. A bibliography doesn't connect any claim to how strong its evidence is.
“ask Claude to extract word-for-word quotes first before performing its task. This grounds its responses in the actual text, reducing hallucinations.”Source: platform.claude.com
Sources
(7)
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Citations return the exact passages that support each claim, so you can verify answers and surface sources to your users.”
↩︎ Why attribution disappears during summarization“When passed back in subsequent conversation turns, cited_text is also not counted toward input tokens.”
↩︎ Why attribution disappears during summarization“the context field is useful for storing document metadata as text or stringified JSON.”
↩︎ Publication and collection dates as required fields“Lists, transcripts, special formatting, more granular citations”
↩︎ Established vs contested sections and fit-for-purpose formats“each text block can contain a claim that Claude is making and a list of citations that support the claim.”
↩︎ Key concept“the context field is useful for storing document metadata as text or stringified JSON.”
↩︎ Exam trap 3“title is limited in length, so the context field is useful for storing document metadata as text or stringified JSON.”
↩︎ Checkpoint - 2.
“This ensures all claims are properly attributed to their sources.”
↩︎ Why attribution disappears during summarization“Each subagent needs an objective, an output format, guidance on the tools and sources to use, and clear task boundaries.”
↩︎ The subagent output contract: claim, source, excerpt“The LeadResearcher synthesizes these results and decides whether more research is needed—if so, it can create additional subagents or refine its strategy.”
↩︎ Conflicting statistics from credible sources“one subagent explored the 2021 automotive chip crisis while 2 others duplicated work investigating current 2025 supply chains”
↩︎ Publication and collection dates as required fields - 3.
“each citation carries the source and title you provided”
↩︎ The subagent output contract: claim, source, excerpt“Any stable string works: a URL, or an internal identifier such as kb://article-1234”
↩︎ The subagent output contract: claim, source, excerpt - 4.
“Type safe: Guaranteed field types and required fields”
↩︎ The subagent output contract: claim, source, excerpt“Type safe: Guaranteed field types and required fields”
↩︎ Publication and collection dates as required fields - 5.https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinationsOfficial docs
“Make Claude's response auditable by having it cite quotes and sources for each of its claims.”
↩︎ The subagent output contract: claim, source, excerpt“For each claim, find a direct quote from the documents that supports it.”
↩︎ Conflicting statistics from credible sources“Explicitly give Claude permission to admit uncertainty.”
↩︎ Conflicting statistics from credible sources“ask Claude to extract word-for-word quotes first before performing its task”
↩︎ Established vs contested sections and fit-for-purpose formats“Make Claude's response auditable by having it cite quotes and sources for each of its claims.”
↩︎ Exam trap 1“For each claim, find a direct quote from the documents that supports it.”
↩︎ Exam trap 2“ask Claude to extract word-for-word quotes first before performing its task”
↩︎ Exam trap 4“Make Claude's response auditable by having it cite quotes and sources for each of its claims.”
↩︎ Prediction“ask Claude to extract word-for-word quotes first before performing its task. This grounds its responses in the actual text, reducing hallucinations.”
↩︎ Checkpoint - 6.https://platform.claude.com/cookbook/misc-using-citationsSecondary source
“The citation feature will not return citations pointing to documents or locations that were not provided as valid sources.”
↩︎ Conflicting statistics from credible sources - 7.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“Encourage source verification: Ask Claude to verify information across multiple sources.”
↩︎ Conflicting statistics from credible sources“wrap each document in <document> tags with <document_content> and <source> (and other metadata) subtags for clarity”
↩︎ Publication and collection dates as required fields