Inngest makes durable agents a serverless execution primitive

Serverless agent products need durable execution without forcing every team to run separate queues, workers, and state machines. Inngest's `llms.txt` describes a durable workflow engine for AI applications with step-level retries, event coordination, throttling, concurrency controls, human-in-the-loop patterns, built-in observability, and no infrastructure to manage. Its Durable Agents documentation is more specific: each LLM call, tool invocation, database write, API request, wait, or sub-agent delegation can be modeled with primitives such as `step.run()`, `step.waitForEvent()`, and `step.invoke()`, so the agent can checkpoint progress, suspend while waiting for a human, and resume without replaying completed work. The same documentation index also exposes Agent Evals, sessions, traces, scoring, flow control, coding-agent CLI patterns, Dev Server MCP, and LLM-ready docs. For agent builders, the practical signal is to evaluate whether a runtime makes dynamic tool loops recoverable, inspectable, rate-controlled, and measurable across production runs instead of treating durability as an external queueing problem.