Five coding agents on one repo is a staffing problem, and worktrees only fix the filesystem

A git worktree per session buys separation of working files and not much past that ^[7][18]. Everything that broke, or nearly broke, sat in a namespace the worktree does not cover ^[18]. Container image tags are global to the registry, so two sessions building different content as 1.0.399 produce a silent overwrite, and the author reports a chained command error that pushed new content onto an old tag, so the behaviour survived the rollback ^[11]. Branch authority is the same class of shared fact: main looked like the base and was not, because production tracked a dedicated deployment branch that main trailed by dozens of versions ^[3].

Read the near-miss as an operations failure rather than an agent failure. Session B's plan was internally coherent, which is to say it announced a frontend build of 1.0.399 from main ^[2], and it was wrong only because of a repository convention that had never been written anywhere the agent could read it ^[13]. Session A caught it, B stopped before pushing and cherry-picked onto the deployment branch instead ^[5], and the human found out afterwards from the logs ^[6]. A correct diff on the wrong base would have taken weeks of shipped features out of production ^[4].

The arithmetic decides how far any of this scales. Coordination is pairwise: three sessions have three channels to keep straight, five have ten ^[17]. Going from three agents to five roughly triples the number of relationships every rule has to hold across, while the cost of adding each agent stays flat. That curve is why human teams grow process, and the recipes here are that process. Territory is assigned by directory, with talk components to the voice session, broadcast to streaming, and callgw to the phone session ^[8]. Shared trees get a hard check of branch and git status before any commit, adopted after one session nearly committed another session's staged changes ^[9].

The item I would lift out of the seven ^[15] is the shape of the stand-up. Sessions declare not only the files they will touch but the ones they will not, and the author credits that second half with nearly eliminating file conflicts ^[10][20]. Declaring intent is cheap. Declaring abstention is what lets a peer keep working without asking permission. It is also the only outcome the post reports, and it is a qualitative one ^[20].

The durable fix is the ledger. Put what is running where into Kubernetes manifests and commit the tag update with each deploy, and any session can read the current state out of the repository instead of querying the cluster ^[14]. A registry check by API before deploying does the same job for tags ^[12]. Both moves convert an ambient fact into a versioned one, which is the only form a peer with no memory of yesterday can be trusted to read. This is one team's operating experience, originally published in Japanese and carried on dev.to as a follow-up to an earlier piece on worktrees ^[16], and it comes with no collision counts on either side of the rules.