Skip to content
DevLaunch home

Guide · Agent orchestration

Run Multi-Agent Research with Sources You Can Check

Assign distinct research questions and require evidence for each claim before a coordinator combines the findings.

By DevLaunchPublished

Several research agents can bring back the same unsupported claim from different websites. Counting the number of workers that agree will not tell you whether the claim is true. A research workflow needs a way to trace each important statement to evidence and distinguish independent confirmation from repetition.

Assign questions, not broad territories

For a hypothetical software-purchase review, ask one worker to verify export options, another to inspect account permissions, and another to identify deployment requirements. Give them the exact product and edition under consideration. A broad assignment such as "research the product" encourages duplicated summaries.

Ask each worker to return a claim, a source link, the relevant passage or section, and the date checked. Keep direct quotations short and separate them from the worker's interpretation. If a source describes a plan or beta feature, that status belongs with the claim.

Keep a claim ledger

The coordinator can maintain a simple list of the findings it plans to use. For each claim, record the evidence, any conflicting source, and whether it is verified, uncertain, or an inference. This makes the synthesis step an inspection of support rather than a contest between polished summaries.

Suppose one worker reports that all data can be exported. Another finds documentation listing only contacts and transactions. Keep the narrower supported statement until the discrepancy is resolved. Do not average the two answers into a vague claim that export is "comprehensive."

Prefer current primary documentation for product capabilities. A blog post can help identify a question to investigate, but it may describe an older version, a different plan, or the author's own setup. The coordinator should preserve those distinctions when writing the final comparison.

Separate facts from recommendations

A recommendation often combines evidence with the customer's priorities. Say which part is judgment. For example, a manual export may be acceptable for an occasional archive but unsuitable for a nightly migration pipeline. The documented capability and the suitability assessment are related, but they are different statements.

Give the coordinator the decision criteria before it reads the findings. Otherwise, it may organize the report around whatever the workers found easiest to describe. A useful purchasing review should answer the buyer's questions even when the evidence for one answer remains incomplete.

Audit the final synthesis

Choose several important claims from the finished report and follow their references back to the source. Check whether the final wording is broader or more certain than the evidence. Also look for findings that disappeared because they complicated the recommendation.

Save unresolved questions alongside the report, with a clear next step for each. A missing answer can become a vendor question or a test in a trial account. It should not become a confident sentence simply because the workflow needs to finish.

Sources & further reading

Keep building

View topic →