The tragedy of unbounded retry

Optimistic clients retry on 429, then retry harder. One runaway job becomes a denial-of-service against your own community.

If docs do not publish the window, operators invent the worst one.

Design the budget in the open

Show remaining budget. Cap concurrency. Prefer queueing over silent failure.

Inside the lab: fuse first, cleverness second.

platform ops diagram
Visual 2/3 — structure for operators

Field steps

  1. Publish soft and hard limits where humans can find them.
  2. Add client-side jittered backoff by default.
  3. Cap concurrent tool calls per run.
  4. Alert on spend spikes before the card declines.

Long view

As AI products commoditize, reliability storytelling beats feature lists. Budgets are how trust shows up in the API.

One client that retries forever can take the edge down for everyone.
Visual 3/3 — takeaway card