Agent Main Thread
Durable, webhook-first single-request inference for patient agent loops.
Agent Main Thread lets an agent submit one inference step, receive a job id immediately, pause its own loop, and resume when BatchRouter delivers the completed artifact callback. The workflow keeps normal batch durability, route scoring, retries, artifacts, and webhook delivery while biasing routing toward the lowest expected wall-clock time under the 24-hour SLA.
A single agent inference step is accepted as durable async work, routed through BatchRouter, and completed by webhook instead of a held HTTP response.
For: Agent platforms, autonomous workers, and orchestrators that can pause between inference steps
What this workflow does
- Return a durable job id immediately so agents never depend on long-lived inference HTTP connections.
- Use telemetry-aware provider scoring, fallback, and retry behavior for single-step agent work.
- Use a bounded adaptive dispatch window before normal work-unit planning when the latency profile allows it.
- Deliver completion through the existing signed webhook and artifact path so agent loops can resume reliably.
Success SLAstandard
Webhook completion inside the 24-hour async inference SLA, with routing biased toward fast eligible lanes.
Defined as: The run reaches a terminal state before the configured standard deadline, writes an artifact for the single request, and delivers the signed webhook callback without requiring the initial REST call to remain open.
Default routing policyAgent main thread
Favor low expected wall-clock time for patient agent loops while keeping cost and 24-hour SLA risk visible.
Ship Agent Main Thread in production
One POST, one webhook. Outcome-priced, with no token math or DAG setup.