Kit · Lab
Retries lab
Retries make an agent run finish. They also make it slower, more expensive and, without idempotency keys, repeat its side effects. Move the sliders and watch all five numbers at once.
Try
Ten steps at a 5% failure rate: without retries, about four runs in ten fail somewhere.
40.1% of runs fail somewhere. Retries would rescue most of them.
- Runs that finish
- 59.9%
- Calls per run
- 10.0
- Expected time
- 17 s
- Worst case
- 50 s
- Duplicate side effects
- none
What it shows
A teaching model, not a benchmark: each call fails independently at the same rate, a failed attempt costs its full timeout, and backoff doubles after every retry. Real dependencies fail in bursts, which makes the worst case more likely than this model says, not less.
Read on: A retried agent job is a second chance to send the same email, or open the Tools component on the map.