Agent runtime
The process and infrastructure that executes an agent's work — starting it, allocating workers, enforcing resource limits, streaming output and terminating it.
Updated 15 September 2026
In practice
A research job runs for forty minutes, pauses two days for an approval, survives a worker crash, and resumes from its last committed step without repeating the payment it already made.
The distinction
The runtime knows where the process is. It does not know whether the business job is complete or whether an external effect committed — those need their own records.
A model call is not a unit of work
A model call is a request and a response. A job has a wider lifecycle: accept an objective and authority, create an identity for the run, execute bounded steps, record checkpoints and effects, wait or retry when required, reconcile external state, deliver and obtain acceptance.
The distinction matters because a model can finish while the job stays unresolved. A tool may time out after the service accepted the request. A person may not answer an approval for two days. A worker may die after a database commit but before recording the response.
Five properties that matter more than the framework
- The run can pause and resume without losing its state.
- A retry is safe because the logical effect has an idempotency rule.
- A worker failure does not silently lose the run.
- Each state transition is inspectable by the people accountable for the job.
- An ambiguous external effect can be reconciled before the run closes.
Framework choice is a smaller decision than it appears. The durable asset is the controlled execution path, and the most valuable control plane is the one that can establish what the outside world did after a timeout.
Sources & context
Temporal documentation
Temporal · 2026
Documents crash-resilient workflows that resume after worker, network or host failure and can run from seconds to years.
Source checked 2026-09-13
Introducing the Agents API
OpenAI · 2026
States that useful agents need a harness managing context, tool use and subagent coordination. A platform claim showing where model providers are expanding.
Source checked 2026-09-13