The repository is the operating manual
People can recover from an undocumented setup step by asking a teammate. An agent cannot. It sees scripts, configuration, tests, and the instructions checked into the repository. When those sources disagree, the agent spends time exploring instead of changing code.
Parallelism exposes weak boundaries
Multiple agents can work quickly only when their ownership is clear. Shared generated files, giant configuration modules, and tests that depend on global state turn parallel execution into merge repair. The practical fix is not more orchestration; it is smaller surfaces with explicit contracts.
Make the happy path boring
A strong agent-ready repository has one setup command, focused test commands, predictable build output, and written rules for dangerous operations. Those conventions help human contributors too. Agent adoption simply makes the cost of their absence impossible to ignore.