Running agents at scale

Recommended

• Give every parallel agent its own working copy

  AID-63

  Run each agent that works at the same time as another in its own worktree or container, on its own branch. Never let two agents edit one checkout.

• Let background agents deliver only pull requests

  AID-64

  An agent that runs without a person watching works under its own bot identity and pushes only to its own branch prefix, such as agent/. It delivers a pull request and never approves, merges or deploys.

• Cap open agent pull requests at what people can review

  AID-65

  Set a cap on open agent pull requests per reviewer, such as 3, and start no new background task for a reviewer at the cap. Close agent pull requests nobody has touched for 7 days.

• Use review bots as a first pass and never as the approval

  AID-66

  Let AI review bots comment on pull requests, but never let a bot's review satisfy a required approval. Measure each month how many of each bot's comments led to a code change.

• Split work between agents by files and merge through one lead

  AID-67

  When several agents work toward one goal, give each a written task that lists the files or folders it may change. No two agents own the same file, and one lead session or person merges the results.

• Attribute the cost of every agent session

  AID-68

  Tag every agent session with the person, the team and the task. Report the cost per team and per merged agent change every month.

• Roll agents out to a team in steps

  AID-69

  Start with a pilot of 2 to 5 people for 4 weeks under a written policy. Widen use only when the change fail rate has not risen from its value before the pilot.