Loop Engineering Concepts › What a turn is, and why you count every one
01 · part 1, conceptsconcept

What a turn is, and why you count every one

One turn is one call to the model, by any agent. Count turns across every agent in the system, or a team of four finishing a job looks exactly like one agent finishing it.

The unit

One turn is one model call by one agent. Not one pass of your loop, not one user request, and not one attempt: one call.

Count turns across every agent, including the ones that never touch the work. A reviewer reading someone else’s answer and an arbitrator settling an argument each spend turns like anyone else.

One turn is one model call, by anybody turns worker 1 6 worker 2 1 worker 3 1 worker 4 1 reviewer 1 never touches the work arbitrator 1 0 1 2 3 4 5 6 7 8 9 10 11 one counter, six callers It does not reset when the acting agent changes. That is the whole reason four agents finishing a job can be compared with one agent finishing it. Total for this configuration: 11 turns across six agents.
The complete loop on the answer job, agent by agent. Read it left to right.
The reviewer and the arbitrator each spend 1 turn without touching the work, out of the 11 this loop takes. Leave them out and the loop looks cheaper than it is, with nothing to warn you.
  • Six agents: four that do the work, a reviewer and an arbitrator.
  • Every call from any of them ticks the same counter.
  • The counter never resets when a different agent acts.

Why the definition comes first

Without it, four agents finishing a job look exactly like one agent finishing it. Both “worked”. The difference is in a number nobody wrote down.

That’s how teams of agents get adopted without evidence: the comparison that would have shown the cost was never made, because the two were never put on the same scale.

what gets comparedwhat it hides
did it get the right answer?how many calls it took to get there
how long did it take?how much of that was one agent waiting for another
how many passes of the loop?four agents inside one pass

Count in one place

A count kept in several places will eventually be wrong, quietly. In the labs, every model call goes through one function, and a separate check confirms the total matches the record of the run. A call that skipped the counter shows up as a mismatch, not as a believable small number.

Do this before adding any layer. Every layer in step 3 is judged by its effect on the count. If the count is unreliable, so is everything you decide about them.

Why the raw count isn’t enough

A loop that takes twenty turns and finishes beats one that takes four and doesn’t. So the count alone can’t rank loops. This volume ranks them by progress per turn: how much of the job got done, divided by how many calls it took. Step 4 measures it, and the result isn’t what the usual advice predicts.

Try it. In the labs folder, run:
python3 lab.py answer

Part 1 lists every turn of the bare loop. The report lists every turn of the complete loop too, including the reviewer’s and the arbitrator’s.