An agent is not a person and it is not a magic box. It is a process that reads a task, writes the tests that prove the task is done, writes the code that passes them, and then stops.
Work arrives already decomposed
Nothing is handed to an agent as "build the booking system". The spec is broken into tasks with declared dependencies, and a task is only dispatched once everything it depends on is finished. That ordering is what lets several agents run at once without treading on each other.
Each task carries only what it needs
An agent receives the instruction, the notes left by the tasks it depends on, a map of the repository, and nothing else. Handing over the whole history would make the cost of a task depend on when it happened to run — and would bury the thing that actually matters.
The pipeline decides what is done
An agent reporting success is a claim, not a verdict. The verdict is the test pipeline going green. If the two disagree, the pipeline wins, and a senior engineer reads the difference.