Running an AI Team: Independent Review After the Role List
We did not mean the same kind of team
I asked for an AI team, but the response described a human team. I asked whether an AI team had actually been implemented and whether supporting documents existed.
Role names and documents still did not convey the workflow I wanted. I clarified that, after an assignment, each member should form an opinion, discuss the work with the others, carry it out, and report back.
“Build a team” bundled several expectations together: separate reviews, comparison of opinions, execution, and a final report. Completing a role document would not establish that this sequence worked.
Independent opinions and cross-review
Asking roles to review independently differs from asking them to compare their conclusions. Separate opinions make disagreements visible. Comparing those opinions then needs to lead to a decision about what to execute.
I subsequently asked for concrete next tasks by role. Later requests also called for independent verification of existing work and written plans and reports.
The table below summarizes the requested workflow, with private task names and business details removed.
| Stage | What I wanted to see |
|---|---|
| Review by role | What each role examined and concluded |
| Comparison | Disagreements and missing checks |
| Execution | Work actually performed after review |
| Verification | Separate checks of existing work and changes |
| Reporting and records | Results and unresolved issues carried into later work |
The session contains actual sub-agent calls and reviews by role, so it went beyond describing a team in documents. Those calls alone do not prove that every task followed the complete sequence consistently.
Repeating the instructions was another problem
Written roles and rules were of limited use if I had to repeat a long instruction to read them on every assignment. I raised that issue too. The rules needed to be read and applied in later work.
This separated two kinds of evidence. A role written down is different from a review performed by that role. A reporting template is different from a report that identifies unresolved problems. One useful report cannot establish that the workflow will persist into the next task.
What the record establishes
This article excludes business data and internal integrations. It does not present a new public demonstration of team performance. The available evidence establishes corrections to the request, actual sub-agent calls, and recorded reviews by role.
There was no comparison against a single AI performing the same task, so the record does not establish better speed or accuracy. The case concerns turning the label “AI team” into specific expectations for independent opinions, execution, verification, and reporting.
This account uses session records from August 11–31, 2026. It describes that historical workflow; no new agent team was run for the article.
A later case, integration bugs missed by individual tests, follows the review roles through to the errors found by checking actual connections.