Sample report · made-up data

Your agents, in numbers.

1,538 steps from 360 runs, Nov 9, 2026 to Nov 15, 2026. Built Nov 16, 2026.

360
Runs
4.2%
Runs that failed
$111
Spend, 7 days
$0.308
Average cost per run
19.3 s
Slowest 5% of runs

Start here

What to fix first

Fixing the items below could save about $245 a month at this volume. These are estimates from your own numbers, not promises.

  1. 1medium≈ $39.21 a month

    19 runs call the same tool 6+ times

    Most often search_orders. That pattern usually means the agent is stuck in a loop or re-asking for data it already has.

    Fix: Add a call limit for search_orders, and give the agent the result of earlier calls so it does not repeat them.

  2. 2medium≈ $27.39 a month

    Retries cost $6.39 in this period

    23 steps needed a retry. Each retry pays for the full call again.

    Fix: Retry only errors that can succeed (rate limits, timeouts), cap retries at two, and fix the step that fails most.

  3. 3medium

    4.2% of runs end in an error

    15 of 360 runs failed. The most common error is "TimeoutError: Timeout after <n>ms calling refund_api" in refund_api (15 times).

    Fix: Start with refund_api: add a retry with backoff or a fallback, and return a clear message to the user when it still fails.

  4. 4medium

    search_orders takes 32% of all step time

    Typical call: 577 ms. Slowest 5%: 7.0 s or more. Users feel the slow tail.

    Fix: Set a timeout near 1.7 s, cache repeat calls, and run it in parallel with other steps when the inputs allow.

  5. 5low≈ $179 a month

    gpt-4.1 is 94% of spend, with short answers

    Its calls average 159 output tokens. Short, structured answers often work on a smaller model.

    Fix: Test a cheaper model on a sample of these calls with the same prompt. Keep gpt-4.1 for the hard cases.

  6. 6low

    15 steps timed out

    Timeouts waste the time the user already waited and often trigger costly retries.

    Fix: Lower timeouts on slow tools and answer from partial data instead of waiting.

Cost

Where the money goes

By model

  • gpt-4.1$104 · 94%

    371 calls · $0.281 each

  • gpt-4.1-mini$6.48 · 6%

    360 calls · $0.018 each

By agent

AgentRunsSpendFailed
support-agent256$78.430.0%
refund-agent104$32.2814.4%

Five most expensive runs

RunAgentStepsCostResult
run_1009support-agent12$0.963Finished
run_1016support-agent11$0.963Finished
run_1029refund-agent14$0.963Finished
run_1082support-agent11$0.963Finished
run_1093support-agent10$0.963Finished

Errors

What breaks, and where

ErrorStepTimesLast seen
TimeoutError: Timeout after <n>ms calling refund_api
e.g. Timeout after 30009ms calling refund_api
refund_api15Nov 15, 2026
ToolFailed: Could not complete refund for order <n>
e.g. Could not complete refund for order 8481
gpt-4.111Nov 15, 2026

15 runs ended on an error. 0 more hit an error but recovered by the end of the run.

Speed

Slow steps

StepCallsTypicalSlowest 5%Share of time
gpt-4.1
model · 11 errors
3712.4 s3.5 s37%
search_orders
tool
703577 ms7.0 s32%
refund_api
tool · 15 errors
104968 ms30.0 s22%
gpt-4.1-mini
model
360612 ms876 ms9%

Trend

Day by day

  • 11-09$16.41 · 52 runs · 4 failed
  • 11-10$15.54 · 52 runs · 3 failed
  • 11-11$14.85 · 52 runs · 2 failed
  • 11-12$16.77 · 52 runs · 2 failed
  • 11-13$16.77 · 52 runs · 2 failed
  • 11-14$15.37 · 52 runs · 1 failed
  • 11-15$15.01 · 48 runs · 1 failed

Want this every day?

Send traces to Uncoil and the report builds itself.

Same numbers, live, with alerts when cost or errors jump. Free up to 5,000 steps a day.

Start free

How we built this: we read your steps once, added up cost, time and errors, and kept only the totals shown here. Prompts, answers and customer data are never stored. A run counts as failed when its last step errored. Savings are rough estimates based on this period only.

Get this for your own agents.