When do multiple agents help?

Add a role only when it solves a problem

Separate workers can help with independent tasks: extracting events from different document batches, checking different issues, or assessing a draft against distinct criteria. They can also repeat the same mistake, lose context between handoffs, or produce incompatible outputs.

Before adding roles, identify the bottleneck. Is it volume, a need for independent criticism, or a missing verification step? If the task is simple, a single controlled workflow can be easier to understand and evaluate.

Define each handoff

For a drafting workflow, the evidence analyst might produce a proposition table; the researcher produces verified authority entries; the drafter receives approved entries; the reviewer identifies unsupported or overstated passages. Each component needs an input specification, output format, completion condition, and exception route.

Do not hand over only a polished narrative. Preserve source IDs, passages, uncertainty, and review status. A shared table with stable fields helps prevent one worker's inference from becoming another's fact.

Work the failure mode

The analyst writes: “Alder likely received the revised schedule.” The drafter shortens it to: “Alder received the revised schedule.” The reviewer sees the analyst's table but not the original and approves the statement.

You have three roles and one unverified inference. Repair the handoff: require the drafter to preserve the proposition's status, and let the reviewer inspect the primary passage and contrary material. Role separation only helps if it changes what is checked.

Use a small review contract

Review the draft against the approved proposition and authority tables. For each material sentence, identify its support and any change in scope or certainty. Read the source passages where supplied. Report unsupported additions, missing qualifications, and omitted contrary material. Do not rewrite the whole draft. Return issues ordered by materiality, with a proposed correction and the source needed to decide it.

The reviewer should not mark its own speculation as a defect that must be adopted. Keep “verified error” separate from “question for counsel.”

Test the whole system

Measure the final output and the intermediate handoffs. If a worker fails or a source is missing, the system should leave an exception rather than a fabricated completion. Compare the multi-worker version with a simpler baseline on the same task. More activity is not necessarily better work.