The Three-Agent Team in the First Recorded Week── Splitting work by role makes it possible to verify each output against a clear criterion

01The operator split work into three roles from the first recorded week

The operator has used generative AI for work since March 2023. The records that survive begin with exchanges with Claude from February 2026 onward. In the very first week of that record, on 8 February 2026, work was already being handed over by role.

The request went like this: create three agents, assign the first to organise requirements, the second to draft an implementation policy, the third to identify test coverage, and then integrate their outputs. A single request contained both the division and the reassembly.

What matters is that this pattern appears at the very start of the record. For the operator, splitting by role was not a new experiment but an established habit. The question is what that habit produces.

02Splitting by role settles the criterion for each output

The immediate effect of the three-role split is verifiability. The requirements output can be checked for completeness. The implementation policy can be checked for consistency with those requirements. The test coverage can be checked for whether it probes the policy's weak points. Each output has a different, specific criterion.

Hand the same work over as a single block and the output comes back as a single block. What to read first, and what standard to judge it against, is left entirely to the person receiving it. When roles are split, those standards are settled before the output arrives, and the effort of checking drops.

03Unverifiable output cannot be judged, regardless of length

The longer an AI's output, the more time it takes to read through. Even after reading, if there is no criterion to judge it against, the conclusion is "it looks plausible." That is a significant loss for the person using AI: time spent without a basis for judgment left behind.

Anthropic's published guidance for Claude Code states that when AI does work, it should be given a check it can run. A test suite, a build exit code, a script that compares output against a reference: anything that returns pass or fail. Without such a check, "looks done" becomes the only signal, and the verification loop falls back to the person.

Splitting by role is one way to create that check on the human side. When each role's output is paired with a criterion, the person can evaluate the AI's work as pass or fail rather than reading for impressions.

04A role, here, is a unit of work whose judgment criterion is singular

The word "role" covers a lot. In this series, a role is a unit of work for which the judgment criterion settles to one clear question. The operator's three roles fit this definition.

When a task with multiple criteria is crammed into a single role, telling which part of the output is good and which is not becomes difficult. Split it down to units where each criterion is clear, and you can mark each output as pass or fail.

The computer scientist Edsger Dijkstra wrote in 1974 that considering each aspect of a problem in isolation, rather than tackling everything at once, is "the only available technique for effective ordering of one's thoughts." The same principle applies to handing work to AI.

05In promotional material review, expression checks and source checks can be split

Consider a case in pharmaceutical promotional material review. When asking AI to do a first pass on an advertising material, one option is to hand it over whole: "review this material." The output will be long, and the reader will not know what standard to apply.

Split by role, the task looks different. The first role checks whether any pre-approval expressions appear. The second checks whether stated figures match the cited source. The third checks whether mandatory disclosures are present. Each role's output can be judged as yes or no.

This division mirrors what reviewers already do by hand. Expression compliance, numerical accuracy against sources, and presence of mandatory items are distinct checks that reviewers perform separately. When the same units are used to split AI work, the outputs can be judged against criteria the team already has.

06Splitting roles separates the maker from the checker

Splitting by role has a second effect: the structure ensures that the output is checked by a role other than the one that produced it.

In the operator's three roles, the test-coverage role inspects the implementation-policy role's output. If the same role that wrote the policy also wrote the tests, it would tend to test only the paths it already considered. A separate role for test coverage can probe the policy's blind spots from the outside.

Anthropic's guidance recommends having a separate agent review a diff against stated criteria, rather than letting the implementing agent grade its own work. A reviewer without the implementer's context judges the result on its own terms, free of assumptions formed during the work. Research by Du et al. in 2023 confirmed that when multiple language-model instances present and debate their answers and reasoning, factual accuracy and reasoning quality both improve.

The same principle exists in human work. In pharmaceutical review, separating the author from the reviewer is a basic quality control practice. When handing work to AI, keeping the creating role and the checking role apart raises the reliability of the output.

07Tomorrow, split work into units with a single criterion each

There are three steps.

1. Count the criteria and create that many roles. Look at the work you want to hand over and count how many things you would check. Each check becomes a role.
2. Write the role and its criterion in one sentence each. "You handle requirements. The output is judged by whether any necessary conditions are missing." One sentence for the role, one for the criterion.
3. Ask for integration last and check for contradictions. Once all role outputs are in, have an integration role check for contradictions. If any are found, redo only the role that caused them.

Anthropic's guidance recommends separating exploration and planning from implementation. Splitting by role follows the same logic: explore, plan, and build are given to different passes so that each can be checked on its own terms. Research on the MetaGPT framework (Hong et al., 2023) reported that assigning specialised roles to agents through standardised procedures and having them verify each other's intermediate outputs reduced cascading errors compared to chaining language models without role separation.