← Notes

03 / Agent roles

Separate the search from the check

The division of work can matter more than the number of agents.

· 2 min read

What DoorDash reports

DoorDash first used specialist reviewers. They found local errors but missed problems across system boundaries.

The next design used two reviewers with broader context. Each had too much to check, and some findings were lost.

The third design added a scout. It identifies possible issues. Two reviewers then investigate those issues. [1]

“The lead scout's job isn't to verify anything.”

DoorDash, on the scout’s role. [1]

Version 1Specialists

Narrow context

Version 2Broad reviewers

Too much to check

Version 3Scout → Reviewers

Find leads, then verify

Our illustration of the three reported designs. The sequence does not represent measured performance.

Our observation

Each role needs a clear result

A possible issue and a verified issue are different results. Separate roles can make that difference explicit.

This design may help when a broad search precedes a detailed check. It may add unnecessary work when one agent can complete both steps.

This is one team’s report. It does not prove that three agents are better than two.

A question for your buildWhat must each agent produce before the next agent can use its result?

Source

  1. DoorDash: How we built an AI code reviewerThe “How we got here” section describes all three designs.