The Agent Is a Workload: What Platform Teams Should Demand Before One Ships
Platform Engineering, Monday, November 9th, 2026
An Amazon Advertising engineer lists what platform teams should require before an AI agent goes to production.
The author, who builds and operates production AI agents, says the asks are the same as for any workload, translated for a system that behaves probabilistically and consumes resources unpredictably.
First is identity: an agent should run as itself with credentials scoped to its job, not through a shared service account or the deployer's permissions.
Second is budgets enforced outside the agent, including rate limits per workflow, a cap on concurrent runs and a spend ceiling that trips automatically.
Agent load is bursty because one question fans out into plans, tool calls and retries, and stop controls and traceable logs complete the list.