CODING CHALLENGE · N°17

Run the Goal Loop

Medium AI EngineeringAgentsLoop Engineering

The goal loop is the engine of loop engineering: attempt the next task, let a verifier judge it, retry failures, and stop when the goal is met — or stop loudly when the cap trips. Implement the whole outer loop as one pure, testable function.

The problem

Implement run_loop(queue, outcomes, max_iterations). Each iteration the agent attempts the task at the front of the queue, and outcomes[i] is the independent verifier’s verdict for iteration i (if the outcomes list is exhausted, the verdict is a failure). A pass moves the task to done; a fail leaves it at the front to retry. The loop ends when the queue is empty or the iteration cap is hit. Return {"done": […], "remaining": […], "iterations": n, "status": "complete" | "cap"}.

EXAMPLE 1

Input queue=["a","b"], outcomes=[True, False, True], max_iterations=10

Output {'done': ['a', 'b'], 'remaining': [], 'iterations': 3, 'status': 'complete'}

"b" fails once and is retried — the verifier decides, not the agent

EXAMPLE 2

Input queue=["a","b"], outcomes=[False, False, False], max_iterations=3

Output {'done': [], 'remaining': ['a', 'b'], 'iterations': 3, 'status': 'cap'}

the cap trips with work remaining — the loop ends loudly, not silently

EXAMPLE 3

Input queue=[], outcomes=[True], max_iterations=5

Output {'done': [], 'remaining': [], 'iterations': 0, 'status': 'complete'}

nothing to do → the goal is already met, zero spend

CONSTRAINTS

One verifier verdict is consumed per iteration: verdict for iteration i is outcomes[i], or a failure once outcomes are exhausted.
A failed task stays at the front of the queue — state untouched, retried next iteration.
Do not mutate the input queue; return a fresh remaining list.
status is "complete" only when the queue emptied; a cap-hit with leftovers is "cap" — even at 0 iterations.

SOLVE IT YOURSELF

Your turn — write it

Edit the stub, hit Run (or ⌘/Ctrl + Enter), and watch the hidden tests. Stuck? the hints are right above and Reveal solution is one click away.

YOUR TASK

Implement run_loop(queue, outcomes, max_iterations) — attempt the front task each iteration, let the verifier verdict decide pass/retry, and stop on goal-met or cap-hit with the full result dict.

HINTS — 5 IDEAS

Copy the queue first — the caller’s list must survive the run.
Loop while the queue is non-empty AND iterations < max_iterations.
The verdict for this iteration is outcomes[iterations] if that index exists, else False — read it before incrementing.
On a pass, pop the front task into done. On a fail, do nothing — the front task is automatically the retry target.
Status at the end: "complete" if the queue is empty, otherwise "cap".

CPython · WebAssembly

Approach, complexity & discussion — open after you solve

The approach

Copy the queue first so the caller’s list survives. Loop while the queue is non-empty and iterations are under the cap. Read this iteration’s verdict before incrementing the counter (an exhausted outcomes list counts as a failure). On a pass, pop the front task into done; on a fail, leave it at the front so it is retried next round. At the end, the status is "complete" only if the queue emptied, otherwise "cap".

Complexity

Time O(max_iterations); space O(n) for the copied queue and the done list.

Common mistakes

Mutating the caller’s queue instead of copying it.
Reading outcomes[iterations] after incrementing — an off-by-one that shifts every verdict.
Reporting "complete" when the cap tripped with work left — a cap-hit with leftovers is "cap", even at 0 iterations.
Forgetting that running out of verdicts means failure, not success.

Where this shows up

This is the outer loop of loop engineering — attempt the next task, let an independent verifier judge it, retry failures, and stop on goal-met or a loud cap. It is distinct from the ReAct inner loop (act/observe), and the same bounded-retry-with-verification shape drives CI reruns, job queues, and any agent that must finish a task list without spinning forever.

Finished this one? 0 / 75 Challenges done

Explore the topic

See this alongside everything else on the same subject — handbooks, system designs, challenges and tools, in one place.

AI Agents & Tools