A Team of Agents - Plan, Code, Review, Test

The problem

A complex task can be split among agents with roles: one plans, one writes code, one reviews it, one tests it. The code that decides who speaks next, what each one sees and when to stop (the orchestration) matters as much as the models. You will write it for four agents. Each is a function from its inputs to a string (in the tests, scripted fakes).

The protocol

plan = planner(task), once.
Then up to max_rounds rounds, starting with feedback = "":
- code = coder(task, plan, feedback);
- review = reviewer(code). If it does not start with "APPROVE", set feedback = "review: " + review and start the next round: no test this round;
- otherwise result = tester(code). If it starts with "PASS", stop: success. Otherwise set feedback = "test: " + result and start the next round.
Return (code, log) on success, or (None, log) if the rounds run out. log is the list of (role, output) pairs in the order the agents spoke: ("planner", plan), ("coder", code), ("reviewer", review), ("tester", result).

Example. The coder writes v1, v2, v3 in turn; the reviewer answers REJECT: no error handling, then APPROVE, then APPROVE; the tester answers FAIL: test_empty, then PASS.

1 — coder gets as feedback: (empty); reviewer: REJECT: no error handling; tester: not run

2 — coder gets as feedback: review: REJECT: no error handling; reviewer: APPROVE; tester: FAIL: test_empty

3 — coder gets as feedback: test: FAIL: test_empty; reviewer: APPROVE; tester: PASS

The result is ("v3", log), where the log has 9 entries: planner, coder, reviewer, coder, reviewer, tester, coder, reviewer, tester.

What to implement

run_team(task, planner, coder, reviewer, tester, max_rounds): return (code, log), or (None, log).

Requirements

The coder must receive the latest feedback, from the reviewer or the tester, prefixed as above.
The tester only runs on approved code.
Count rounds by calls to coder: it is called at most max_rounds times.