Agent Cost Runaway Checklist

Self-check score

Looks calm — keep it that way

0/12

1. You don't know today's Claude Code / Cursor spendNo dashboard, no daily rollup — only surprise invoices.
2. No daily or monthly budget cap is configuredAgents keep calling until the provider key dies or finance notices.
3. Nothing returns HTTP 429 / stops the loop when over budgetWarnings without enforcement still burn money overnight.
4. Spend isn't tagged by project / agent / tagYou can't tell which workflow is the expensive one.
5. Failed tool calls silently retry expensive modelsOne stuck loop can multiply cost by 10× in an hour.
6. Default model is a frontier model for every tiny taskNano/Haiku/Flash tiers exist for a reason.
7. Agents never estimate cost before a big runPreflight would catch '$12 for this refactor' before it starts.
8. Long agent loops have no should-continue gateBudget health should decide continue / slow_down / stop.
9. One shared API key across all agents and humansBlast radius is the whole company credit card.
10. No alert when today's spend spikes vs the 7-day averageAnomalies are usually runaway agents, not intentional work.
11. You only discover cost after the provider bill arrivesMetadata-first gateways catch it at request time.
12. Prompt caching / cheaper models aren't part of the playbookSame system prompt rewritten every turn is free money burned.

AI Cost Gate's cost-status, preflight, and should-continue turn this checklist into runtime gates — not another doc nobody re-reads.

Occasional emails when prices move or new ACG tools ship.

Official list prices, sorted by value

How much will your agents cost this month?

Paste a prompt, see the bill

Make coding agents budget-aware