- In short
- Attribution preservation is the design property of a multi-agent pipeline in which every agent, retriever, subagent, and synthesiser, carries claim-source mappings forward instead of discarding them. When any intermediate agent strips attribution while processing, the provenance chain breaks there, so diagnosing preservation means tracing where in the pipeline the mapping was dropped.
What attribution preservation demands of a pipeline
Attribution preservation is the analyse-level cousin of the structured claim source mapping. The earlier knowledge point taught you to bind each claim to its evidence at the moment of retrieval. This one asks a harder question: does that binding survive the journey through a multi-agent system, where the finding is handled, condensed, and merged by several agents before anyone sees the output? A mapping that is perfect at the retriever and absent at the final report has not been preserved, and the whole point of the exercise is lost.
The defining property is uncompromising. Each agent in the pipeline must preserve claim-source mappings; synthesis agents must merge mappings from multiple sources; and the final output must retain the full provenance chain. Preservation is not something the last agent can add back at the end, because by then the links are gone. It is a constraint every hop has to honour, which is what makes it a design property of the pipeline rather than a feature of any single agent.
- Attribution preservation
- A property of a multi-agent pipeline in which claim-source mappings are carried forward by every agent, retrievers, subagents, and synthesisers, so that the final output retains a complete provenance chain from source to claim.
The chain breaks at a single hop
The reason this is an analyse-level skill is that the failure is non-local. The symptom appears at the end, a final report whose claims cannot be traced, but the cause is upstream, at whatever agent first rewrote content without forwarding the mapping. A pipeline can have nine agents that faithfully preserve attribution and one intermediate summariser that condenses findings into clean prose and silently discards the structured provenance, and the result is a total loss of traceability that looks, superficially, like the final agent's fault.
This is why diagnosing attribution loss is a tracing exercise. You walk the pipeline hop by hop and ask, at each one, did the mapping go in and did it come out? The break is at the first agent where the answer flips from yes to no. The connection to structured context passing and to coordinator responsibilities is direct: the coordinator defines the contract that each agent's input and output must satisfy, so a systemic attribution failure is traced back to the coordinator that failed to require the mapping, not to whichever subagent happened to drop it.
How the pipeline carries provenance
The healthy path keeps the mapping attached from retriever to subagents to synthesiser to output. The dashed path shows the failure mode: one subagent processes its findings into prose and forwards only the prose, and from that point on the claims it handled are untraceable no matter how careful every later agent is. Anthropic's own multi-agent research system addresses this by separating concerns, subagents gather, the lead agent synthesises, and a dedicated citation stage attaches sources to the final claims, so that attribution is an explicit, owned step rather than something each agent is trusted to remember.
Worked example: a due-diligence pipeline that loses its sources
Worked example
A due-diligence system uses a coordinator, two research subagents, an intermediate summariser, and a final report writer. The finished report makes strong claims but cites nothing, even though the subagents retrieved well-attributed passages.
To analyse this, you trace the pipeline rather than rewrite the final report. The two research subagents are checked first: their outputs still carry the claim-source mappings, with excerpts and dates intact. So the break is not at retrieval. The final report writer is checked next: it faithfully renders whatever it receives, and it received no mappings. So the writer is not the culprit either. It cannot cite sources it was never given.
The break is the intermediate summariser. It was prompted to condense the subagents' findings into a tight narrative, and it did exactly that, by collapsing each finding into a sentence and discarding the structured provenance as if it were noise. Every claim that passed through it lost its source there. The fix is not to patch the report writer or to add a disclaimer; it is to change the summariser's contract so it merges and forwards mappings instead of dropping them, and ultimately to make the coordinator require a structured, provenance-carrying format from every agent.
The lesson the exam wants is the diagnostic one: the symptom surfaced at the output, but the root cause was a specific upstream hop, and the durable fix lives at the coordinator that owns the inter-agent contract.
Common misreadings to avoid
Misconception
The final report writer is responsible for citations, so an uncited report is a bug in the last agent.
What's actually true
Misconception
As long as each agent works correctly on its own, attribution will naturally survive the pipeline.
What's actually true
Why structured passing beats inline prose
The deeper reason attribution gets lost is that prose is a lossy carrier for structured information. When a finding's provenance lives as an inline phrase, as one source notes, according to a recent report, the next agent that summarises has every incentive to treat that phrase as filler and cut it, because it reads as throat-clearing rather than content. Inline attribution survives only as long as no one compresses it, which in a multi-agent pipeline is never. This is the precise problem structured context passing is meant to solve: provenance carried as a discrete field travels intact because it is data the next agent must forward, not prose it is free to trim.
The practical implication is that the format of the hand-off between agents is the real lever for preservation. If subagents return findings as plain summaries, attribution is already dying at the first hop, no matter how careful the synthesiser is. If subagents return findings as structured records, claim, source, excerpt, date, the synthesiser has something it can merge and forward without reconstructing anything. Preserving attribution is therefore less about exhorting each agent to remember its sources and more about designing the inter-agent message format so that forwarding provenance is the easy default and dropping it takes deliberate effort.
Merging mappings without collapsing them
Synthesis agents have a harder job than retrievers, because they do not just carry mappings forward, they combine them. When a synthesiser draws a single conclusion from three subagents' findings, the resulting claim should carry all three sources, not the one the synthesiser happened to weight most heavily. Merging means union, not selection: the provenance of a synthesised claim is the provenance of every input that supports it. An agent that quietly keeps only the strongest source has thinned the provenance chain even though it never fully broke it, and a reader who later distrusts that one source has no idea two others agreed.
This is also where conflict handling and preservation interlock. If two of the three subagents disagree, a faithful synthesiser does not preserve attribution by picking a side; it carries both values with their separate sources, exactly as conflict handling in synthesis requires. Preservation and conflict handling are two views of the same discipline: keep everything the sources said, bound to who said it, all the way to the output. An evaluator analysing a pipeline checks not only that some attribution survived but that the right, complete set of attributions survived each merge.
Designing for traceability from the start
Because the failure is non-local and the fix lives in the contract, attribution preservation is far cheaper to design in than to retrofit. A pipeline built from the beginning around a provenance-carrying record, every agent accepts it, every agent emits it, preserves attribution almost for free, because no single hop has to be heroic. A pipeline that treated provenance as an afterthought has to be traced, patched at the offending hop, and then re-verified end to end, and even then a later change to any agent can silently reopen the break. The coordinator that owns the inter-agent contract is the right place to make traceability a precondition rather than a hope, which is why systemic attribution failures are correctly attributed to it.
A quick audit you can run on any pipeline
Because the defect is non-local, it pays to have a fixed procedure for finding it rather than staring at the final output hoping the break reveals itself. The audit is mechanical. Start at the source and walk forward one agent at a time, and at each agent ask two questions: did a complete claim-source mapping arrive on its input, and did a complete mapping leave on its output? As long as both answers are yes, the agent is innocent and you move on. The first agent where a mapping arrives but does not leave is the break, and it is the only place a fix will actually help.
This forward walk is more reliable than reasoning backward from the symptom, because the symptom, an unsourced final report, is consistent with a break at any upstream hop and points at none of them in particular. It also disciplines you against the two most common misdiagnoses: blaming the final writer, which the walk exonerates the moment you confirm it received nothing to cite, and blaming the retriever, which the walk exonerates the moment you confirm its outputs were well-attributed. Once the offending hop is located, the durable remedy is to amend the contract that hop operates under, and then to push that same provenance-carrying contract up to the coordinator so the break cannot quietly reopen elsewhere.
How this shows up on the exam
This is one of the more demanding knowledge points in Domain 5 because it asks you to reason about a system rather than a single response. The exam will present a multi-agent pipeline whose final output has lost its sources and ask you to locate the failure or choose the durable fix. The strongest answers trace the break to a specific intermediate agent that stripped the mapping and assign the systemic fix to the coordinator that should have required a provenance-carrying format from every agent. Weaker answers blame the final writer or propose bolting citations on at the end, both of which miss that the links were already gone. Master this and you are ready for the evaluative knowledge point, where you judge a complete synthesis against every 5.6 criterion at once.
In a multi-agent research pipeline, subagents return well-attributed findings but the final report cites no sources. An intermediate summariser sits between the subagents and the writer. What is the best analysis and fix?
People also ask
How is source attribution preserved across multiple agents?
Where does attribution get lost in a multi-agent pipeline?
Whose job is it to keep provenance in a multi-agent system?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
Anthropic: How to Build Multi Agent Systems
Why watch: Walks through Anthropic's orchestrator-subagent research architecture including the dedicated citation pass, showing how claim-source mappings are preserved as findings flow from subagents through synthesis to the final report.
More videos for this concept
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.