Claim-Source Mappings: Keeping Provenance Through Multi-Agent Synthesis

1.Where attribution gets lost

A multi-agent research system is built to compress. Anthropic describes its Research feature this way: "The essence of search is compression: distilling insights from a vast corpus." Subagents search in parallel, each in its own context window, and then condense the most important tokens for the lead agent. Compression is the reason the architecture works. It is also where provenance disappears.

The Agent SDK documentation makes the isolation explicit: "The parent receives a concise summary, not every file the subagent read." That is good for the parent's context budget. But the subagent's final message is the only record of what it found. If that message says "the market grew 14%" without naming the page, excerpt, or report behind the number, the claim-source link is gone. No later step can rebuild it, because the evidence stayed in a context that has already been thrown away.

The same loss happens at every handoff. Anthropic's guidance on multi-agent systems describes teams who "suffered from lost context at each handoff". It names "summarizing results for handoffs" as one source of overhead. Every summarization step tends to keep the gist and drop the detail, and a citation is detail.

Context compaction inside one long-running agent works the same way. In Anthropic's context-engineering cookbook, a compaction summary kept the high-level facts but dropped "three OBSCURE specifics (appendix table cells, unlikely to make the cut)". Source mappings are the kind of specific a summarizer treats as low priority. So if provenance matters, it has to be carried in a structure the summarizer is told to keep, not left to survive by chance.

Sources1234

2.The subagent output contract: claim, source, excerpt

The fix starts with the subagent's instructions. Anthropic's lesson from building Research is that "Each subagent needs an objective, an output format, guidance on the tools and sources to use, and clear task boundaries." For synthesis work, the output format is where provenance is either kept or lost. Don't ask for "a summary of findings". Ask for a list of claims where each claim has its source URL or document name and the excerpt that supports it.

The prompting guidance supports the excerpt requirement. For long documents, ask Claude "to extract word-for-word quotes first before performing its task". The hallucination guide adds that you can make a response auditable "by having it cite quotes and sources for each of its claims." A quote plus its source is exactly the record a downstream agent needs to keep.

The Citations API shows what one such record looks like. Each citation pairs the cited text with the document it came from and the location inside that document:

{
  "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
}

You can carry document metadata the same way when you feed sources in. The prompting guide recommends wrapping each document "with <document_content> and <source> (and other metadata) subtags". In the Citations API, the context field is described as "useful for storing document metadata as text or stringified JSON."

Prompting can still leave gaps: a subagent may skip the source on one claim. Structured outputs close that gap. They constrain the response to a schema, with "Guaranteed field types and required fields". If source and excerpt are required fields on each claim object, no claim can come back without them.

{
  "type": "document",
  "source": {
    "type": "text",
    "media_type": "text/plain",
    "data": "Plain text content..."
  },
  "title": "Document Title",
  " ? ": "Context about the document that will not be cited from",
  "citations": { "enabled": true }
}

Sources25678

3.Preserving and merging mappings in the coordinator

Structured subagent output is only half the job. The coordinator then has to keep those mappings when it combines them. Anthropic's Research system shows the pattern. Subagents return findings, and "The LeadResearcher synthesizes these results and decides whether more research is needed". When the research loop ends, the system "passes all findings to a CitationAgent, which processes the documents and research report to identify specific locations for citations." The stated goal: "This ensures all claims are properly attributed to their sources."

That design hands both the findings and the source documents to a dedicated attribution step. It doesn't hope that attribution survives the synthesis prose. In your own pipeline, the coordinator should treat claim-source records as data to merge, not text to rewrite.

When two subagents report the same fact with different wording and different sources, the safe merge keeps one claim and attaches every supporting source to it. Picking one subagent's citations throws provenance away. When a subagent finds conflicting values, it should finish its analysis and return both values with their sources and a note on the conflict. Deciding how to reconcile them belongs to the coordinator, which sees all the subagents' results, before anything goes to final synthesis.

Citations add one more guarantee. Anthropic's citations cookbook notes that the feature "will not return citations pointing to documents or locations that were not provided as valid sources." A built-in citation step can't invent a source the pipeline never had.

Sources279

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.If a subagent read and cited real sources while it worked, the coordinator can always go back and recover those sources later.Why is that wrong?

   Intermediate tool calls and results never leave the subagent. If the final message doesn't carry the claim-source mapping, the provenance is gone.

   Covered in Where attribution gets lost

2. 2.Telling a subagent to 'include sources where relevant' is enough to guarantee every claim comes back with one.Why is that wrong?

   Prompting alone can still produce missing fields. Making source and excerpt required fields in a structured-output schema guarantees they are present.

   Covered in The subagent output contract: claim, source, excerpt