LoopLlama bills by the workflow step. Usage-based pricing is only fair if the unit is something you can predict and reason about, so it's worth being precise about what a step is — and, just as importantly, what isn't.
A step is one agent turn#
A step is a single agent taking a single turn in a run: it receives the accumulated context, reasons, optionally calls tools, and produces an output. A crew of four agents that each take one turn is four steps. If a workflow loops an agent back for a second pass, that's another step. The mental model is simple: count the turns.
Tool calls made within a turn don't each count as a step. An agent that calls three tools while producing its output is still one step — the tools are part of how that agent does its turn, not separate billable units.
What doesn't count#
Some things happen during a run that you shouldn't pay for, and we don't meter them:
- Retries caused by our own infrastructure — a transient provider error we recover from is on us, not you.
- Time a run spends paused awaiting human input or approval. Waiting is free.
- Queueing and scheduling overhead between steps.
Keeping runs predictable#
Because steps are countable, you can cap them. Every run accepts a budget — a maximum number of steps and a maximum number of tokens. When a run hits its budget it stops cleanly rather than spiraling, which protects you from a misbehaving prompt that would otherwise loop indefinitely.
Pair budgets with the per-step traces every run produces and cost stops being a mystery: you can see exactly which agents consumed which steps, set limits accordingly, and route cheap work to cheap models. Honest metering and good observability are the same project from two angles.