How we cut our agent's tool-call loop from 9 turns to 3

Point 2 is the one nobody talks about. Every 'check then act' pair in my tools was a retry loop waiting to happen.

Would love the eval harness. Especially how you count turns when the model calls two tools in parallel.

@edgar parallel calls in one assistant message count as one turn for us — we measure round-trips, because that is what costs latency. I will clean up the harness and post it in Community.

Did trimming tool output change accuracy at all? I worry about the model missing a field it needed.

Did trimming tool output change accuracy at all? I worry about the model missing a field it needed.

Accuracy went up slightly (94.1% → 95.3% on our 400-booking set). Less noise to read through.

Saving this thread. The budget-in-the-prompt trick feels too simple to work and yet.

We run an agent that books freight: it checks rates, holds a slot, confirms with the customer. Early versions took 9 tool turns on average for a booking. Here is what got us to 3.

1. Return less from tools

Our search_rates tool returned the full carrier payload — 14 KB of JSON. The model spent turns re-asking for fields it had already received but could not find. We now return a 6-field summary and a detail_id for anything else.

2. Make tools do one thing completely

hold_slot used to require a prior check_availability call. The model often skipped it, got an error, then retried. We merged them: hold_slot checks availability itself.

3. Put the plan in the system prompt

You will do at most three tool calls: search, hold, confirm. Do not call search twice.

Surprisingly effective. Explicit budgets beat clever reasoning.

Happy to share the eval harness if people are interested.