Agent tool-loop retry budget: worksheet + mock 429/503 injection (API $0)
Canonical: Agent tool retry budget worksheet on eWorld.live
Check date: 2026-09-28 PT · API spend $0 (local mock HTTP)
Printable retry budget + author failure-injection table on a local mock (get_weather vs charge_once) — not an OpenAI rate-limit page paraphrase. Real live OpenAI hits stay KEY-GATED / unused.
Desk defaults (excerpt)
| scenario | max_attempts | max_total_wait_s | nested SDK retry? |
|---|---|---|---|
| Idempotent read (temporary 429/503) | 4 | 30 | Disable if app backs off |
Side-effect write (charge_once) |
1 (or 2 w/ idempotency key) | 10 | Disable |
401 / credit_balance_exhausted |
1 | 0 | N/A — retry won’t help |
Injection row highlight: nested “SDK×app” path turned 1 logical tool into 3 HTTP on a sticky slow_down — measure http_calls_per_logical_tool before you ship.
Worksheet + Table B cases:
https://eworld.live/news/2026/09/agent-tool-retry-budget-rate-limit-worksheet/
Disclosure: I help run eWorld.live.