- In short
- Handling conflicting sources is the practice of surfacing disagreement between credible sources instead of resolving it silently. When two reputable sources report different figures, the agent annotates both values with their attribution and lets the human consumer judge, rather than arbitrarily selecting one and hiding the conflict.
Why handling conflicting sources is a design decision
Handling conflicting sources is where multi-source synthesis stops being a summarisation task and becomes a judgement task. The moment two reputable sources report different numbers for the same quantity, your agent faces a fork: it can quietly choose one figure and present a clean, confident answer, or it can report that the sources disagree and show both. The first path looks tidier. The second is the one the exam rewards, because a hidden conflict is a defect that surfaces later at the worst possible time, in front of the consumer who acted on a number that was never as settled as it looked.
The principle is deliberately strict. Do not arbitrarily select one value. When credible sources conflict, annotate both with their attribution and let the consumer decide. The word arbitrarily is doing real work here: it rules out the lazy heuristics, take the bigger number, take the newer report, take the first one retrieved, that feel like resolution but are really just concealment dressed up as a decision.
- Conflict handling in synthesis
- The practice of detecting when two credible sources disagree and surfacing that disagreement, both values, each with its own attribution, rather than silently selecting one and presenting a false consensus.
The conflict is the finding
It is tempting to treat a disagreement as a problem to be eliminated before the report goes out. Reframe it: the disagreement is a result. If two well-regarded analyst firms put a company's revenue at different figures, the fact that credible estimates diverge is exactly the kind of signal a decision-maker needs. Flattening that into a single number does not give them a better answer; it gives them a more confident wrong impression of certainty.
This is why conflict handling depends on the structured claim source mapping you learned first. You can only annotate both values honestly if each value already carries its source, excerpt, and date. Without that mapping, two conflicting numbers are just two numbers, and the cheapest way to reconcile them is to drop one. With the mapping intact, the conflict becomes presentable: here is value A from this dated source, here is value B from that dated source, and they do not agree.
Detecting versus resolving
A clean way to think about the workflow is to separate detection from resolution and notice that the agent's job ends earlier than people expect.
The detection step asks a narrow question: do the values match? The interesting branch is the no branch. Before declaring a contradiction, a careful agent checks whether the gap has an innocent explanation, different measurement dates or different definitions, which is the bridge to temporal awareness in source data. If no such explanation exists, the agent annotates both values and stops. Choosing between two genuinely conflicting credible figures is a consumer's prerogative, not the agent's.
Worked example: a competitive-intelligence agent with two revenue figures
Worked example
A competitive-intelligence agent compiles a briefing on a rival's annual revenue. Two credible analyst reports give materially different figures for the same fiscal year.
The agent retrieves both reports and builds a finding for each: one analyst puts revenue at roughly four point two billion, the other at roughly four point eight billion, both for the same fiscal year, each with its source and excerpt captured. A naive agent now picks one, perhaps the higher figure, perhaps the report it happened to read last, and writes a confident single sentence.
That is the trap. The right move is to first check whether the gap is explainable: are the firms using the same fiscal-year boundary and the same definition of revenue? Here they are, so the disagreement is real. The agent therefore writes the briefing to say that credible estimates diverge, presents both figures side by side with their attributions, and notes the gap explicitly. The decision-maker reading the briefing now knows the number is contested and can apply their own risk tolerance, which is precisely the outcome a hidden single figure would have denied them.
Notice what the agent did not do. It did not average the two figures into a fictitious four point five billion that neither source supports, and it did not bury the conflict in a footnote. Surfacing the disagreement with full attribution is the deliverable.
Common misreadings to avoid
Misconception
When two sources conflict, I should pick the more recent one, since newer data is more accurate.
What's actually true
Misconception
Presenting two different numbers makes the report look indecisive, so I should resolve it to one clean figure.
What's actually true
The heuristics that feel like resolution but are not
Most wrong answers to a conflict are not reckless; they are reasonable-sounding shortcuts that an agent reaches for because doing nothing feels like a failure to be helpful. It is worth naming them, because the exam builds distractors out of exactly these moves. Take the more recent source assumes recency equals accuracy, which is false for revised figures, preliminary estimates, and anything seasonal. Take the higher number is sometimes dressed up as conservative planning, but it is just a bias with a justification stapled on. Take the first one retrieved is the laziest of all, because the order of retrieval has nothing to do with which figure is right, yet it is the default an unguided agent falls into.
What unites these heuristics is that each one makes a decision the agent was never authorised to make and then hides that it made it. A planner who later relies on the chosen figure has no idea a second credible estimate existed, so they cannot weigh it, question it, or escalate it. The shortcut did not resolve the uncertainty; it transferred the uncertainty to someone who can no longer see it. That is why the standard is phrased as a prohibition on arbitrary selection rather than a preference for annotation: the failure is the concealment, and the annotation is simply the honest alternative.
Annotating a conflict in practice
Annotation is not just printing two numbers next to each other; done well it preserves enough structure for the consumer to act. A good annotation keeps each value bound to its own source, excerpt, and date, states explicitly that the sources disagree, and, where possible, gives the reader the handful of facts they would need to adjudicate, who published each figure, how recent each is, and whether their definitions match. The reader should come away knowing not just that two numbers exist but why a reasonable person might trust one over the other, without the agent having made that call for them.
It also helps to distinguish the magnitude of the disagreement. Two estimates that differ by a rounding error rarely warrant the same prominent treatment as two that differ by a third, and a thoughtful synthesis can note a minor spread in passing while flagging a material divergence prominently. None of this crosses the line into resolving the conflict; it simply renders the conflict legibly. The agent's contribution is clarity about the disagreement, which is a genuine analytical service, as opposed to the false service of pretending the disagreement is not there.
Why surfacing conflict builds rather than erodes trust
There is a natural worry that showing two numbers makes an agent look unsure and therefore less authoritative. In practice the opposite holds for the audiences that matter. A decision-maker who later discovers that an agent silently dropped a credible competing figure stops trusting every clean number the agent ever produced, because they now know the cleanliness might be concealment. An agent that surfaces disagreement when it exists earns the right to be believed when it presents a single figure, because the single figure now means the sources actually agreed.
This is the same instinct Anthropic describes in its multi-agent research system, where a lead agent synthesises the findings of parallel subagents and a dedicated citation stage ties claims back to sources rather than smoothing them into an unsourced consensus. The synthesis is trusted precisely because it does not manufacture agreement that the underlying evidence did not contain. For the exam, internalise the trade: a little surface-level tidiness is exchanged for durable credibility, and that is always the right trade when a real decision rests on the number.
Annotating a conflict is not the same as reporting everything
A fair objection to the annotate-both-values rule is that it could degenerate into an agent that refuses to take any position and simply dumps every source it found onto the reader. That is not what conflict handling asks for, and the distinction matters. The rule applies to genuine disagreements between credible sources on the same quantity, not to every minor variation or every source of differing quality. An agent should still discard sources that are not credible, still reconcile differences that a date or definition explains, and still draw the conclusions the evidence actually supports. What it must not do is collapse a real, unresolved disagreement between trustworthy sources into a single number as if the disagreement never existed.
In other words, conflict handling narrows the agent's discretion in exactly one situation and leaves it intact everywhere else. The agent remains an analyst, not a passive pipe; it filters, it explains, it synthesises. It simply stops short of the one move it has no authority to make, privately choosing between two credible, conflicting figures and presenting the choice as settled fact. Keeping that boundary sharp is what prevents the rule from being caricatured into indecision, and it is the version of the principle the exam expects you to apply.
Make conflict-checking an explicit workflow step
Anthropic's guidance for building research and synthesis skills treats conflict handling as a named step in the workflow rather than something the model is trusted to do incidentally. The recommended flow is to read all the source documents, identify the key themes, cross-reference each major claim against the material that supports it, assemble a structured summary, and then verify the citations before the output is considered finished. Conflict detection lives inside the cross-referencing step: when you check that every claim actually appears in a source, the cases where two sources say different things fall out naturally instead of slipping past unnoticed.
Building the check in this way matters because an unguided pass tends to produce a fluent answer first and look for support afterwards, which is exactly when a disagreement gets smoothed over. Making cross-referencing and citation verification explicit, ideally as a checklist the agent must complete, forces it to confront each claim against its evidence and to record where the sources diverge. For a customer-facing synthesis especially, that discipline is what turns the annotate-both-values rule from a good intention into a step that actually runs.
It is also why Anthropic's larger research systems separate gathering from synthesis and give citation its own dedicated stage: a single unstructured pass has no natural place to surface a conflict, whereas a workflow with explicit cross-reference and verification steps does. The lesson for the exam is that conflict handling is not only a rule about what to output; it is a workflow you design so that the conflict is detected in the first place.
How this shows up on the exam
Task Statement 5.6 leans heavily on this knowledge point because it is where good intentions go wrong. The exam will describe an agent that confidently reports a single statistic, then reveal that a second credible source disagreed, and ask what the agent should have done. The strongest answer is always to annotate both values with their attribution and let the consumer decide. Distractors will offer plausible-sounding heuristics, take the newer source, take the higher number, average them, and each one is wrong for the same reason: it resolves a conflict the agent had no authority to resolve and erases information the consumer was entitled to see. Pair this with conflict handling's prerequisite, the claim source mapping, and you have the core of provenance-aware synthesis.
A market-research agent finds two credible firms reporting different market-size figures for the same year, with the same definitions and scope. How should the agent present this in its synthesis?
People also ask
What should an agent do when two sources disagree?
Is it ever right to just pick one source?
How is conflict handling different from temporal awareness?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
How Anthropic built their multi-agent research system?
Why watch: Walks through Anthropic's orchestrator-worker research system where the lead agent synthesizes subagent findings, identifying where sources agree versus contradict rather than arbitrarily picking one value.
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.