AI Agent Orchestration Platform · View 17 of 32 · 4 · Runtime
Decisions
- Admission is where a run can be refused: over budget, policy denied or a pinned version that no longer resolves
- Tokens stream to the caller from the gateway while the run is open, so perceived latency is model latency, not orchestration latency
- The result is not final until the completion checkpoint commits — a response acknowledged before that would be a lie under crash
Timing budget
- Edge to admission decision p95 under 60 ms; admission to first dispatch p95 under 60 ms
- First token to the caller p95 under 900 ms end to end for an interactive agent
- Tool hop budget 5 s default, declared per tool and enforced by the executor
Not shown
- The asynchronous submission path, where the caller receives an execution id and polls or subscribes
- Multi-agent delegation — see Multi-Agent Coordination
- Failure and retry — see Durable Execution