Agent-run workstation
Configure type = AGENT_RUN for agent worker dispatch through an AGENT_WORKER backend.
How To Use
Author a workstation with type AGENT_RUN and bind an AGENT_WORKER. Choose behavior for the scheduling axis you need — STANDARD for fire-once dispatch, REPEATER for change-triggered loops, CRON for schedules, or POLLER for long-lived poller ingress. Wire inputs plus outputs, onContinue, onRejection, and onFailure so agent outcomes route correctly. Do not set operation, operationBindings, classificationRoutes, promptFile, or cron fields that belong to other WorkstationType or WorkstationKind families. Choose INFERENCE_RUN for bounded inference, SCRIPT_RUN for scripts, or MODEL_INVOKE for provider-neutral operation bindings. AGENT_RUN is not a scheduling behavior and is not MODEL_INVOKE — keep operation and operationBindings on the model-invoke type, and keep one-shot inference on INFERENCE_RUN.Discriminator: type = AGENT_RUN
AGENT_RUN requires an AGENT_WORKER companion and is compatible with every published WorkstationKind. Bind the agent worker, then choose scheduling on the behavior axis.
Schema reference
The embed below shows the live Workstation base definition with AGENT_RUN overlay applicability. AGENT_RUN carries no field of its own — it is identified by its discriminator and its required AGENT_WORKER companion, and shares the identity and dispatch fields with every other WorkstationType. Excluded fields belong to other WorkstationType families or to CRON scheduling and must not appear on this type. Open the full Factory schema reference for exhaustive field lookup rather than copying contract text into this page.Workstation
objectA processing step in the factory graph. Workstations consume authored work states, run a worker or logical move, and emit the next work states.
- additionalProperties
false (closed)
Fields
Scheduling behavior for this workstation, such as STANDARD, REPEATER, or CRON execution.
- body
bodyOptionalstringInline workstation instructions or script body when authored directly in factory config.
Explicit label-to-destination routing used only by CLASSIFIER_WORKSTATION definitions. Each route must declare a unique non-empty label and one or more outputs.
- copyReferencedScripts
copyReferencedScriptsOptionalbooleanCopy supported referenced script files into the expanded workstation layout when config expand runs.
Cron trigger configuration for workstations whose behavior is CRON.
Optional localized customer-facing explanation of this workstation.
Environment variables added to the workstation execution context.
Guarded loop breakers should use `VISIT_COUNT` guards here with a `LOGICAL_MOVE` workstation instead of top-level exhaustion rules.
- id
idOptionalstringOptional durable public identifier for this workstation. Graph and layout references should use this id instead of the mutable name.
Work states this workstation can consume before it dispatches.
Retry and execution ceilings applied to this workstation.
- name
nameRequiredstringCustomer-authored workstation name used by guards, diagnostics, and authored references.
Optional destination emitted when the workstation makes partial progress and should continue iterating. Classifier workstations must not declare onContinue.
Optional destination emitted when the workstation fails permanently.
Optional destination emitted when the worker rejects the current work without a hard failure. Classifier workstations must not declare onRejection.
Uppercase provider-agnostic operation requested by `MODEL_INVOKE` workstations, such as `TTS`.
Optional workstation-authored slot bindings that resolve operation inputs from runtime content or static config content.
Optional worker-output parsing mode for model workstations. When set to `decision-envelope`, agent output is parsed as a reviewer/checker JSON envelope that maps directly onto WorkResult outcome, feedback, output, and optional recorded output work instead of stop-token routing.
- outputSchema
outputSchemaOptionalstringExcludedJSON schema string used to validate or parse structured model output when configured.
Work states emitted after a non-classifier workstation succeeds. Classifier workstations must use classificationRoutes instead of normal success outputs.
- promptFile
promptFileOptionalstringExcludedPath to a prompt template file loaded for model-oriented workstation execution.
Resource capacity this workstation consumes while one dispatch is in flight.
Optional workstation-specific runner override. When omitted, dispatch falls back to the factory runner, then worker modelProvider compatibility when no explicit runner is configured, then the default codex runner.
- stopWords
stopWordsOptionalarrayExcludedStop words authored on the topology entry for model-oriented dispatches.
Runtime workstation implementation type, equivalent to the workstation AGENTS.md frontmatter type.
Optional policy for whether downstream work uses the workstation output payload or preserves the consumed input payload.
- worker
workerRequiredstringName of a worker declared in the workers list.
- workingDirectory
workingDirectoryOptionalstringGo template resolved from token tags at dispatch time.
- worktree
worktreeOptionalstringGo template resolved and passed as the worktree path to CLI dispatchers.
Examples
Use the minimal valid example when starting an AGENT_RUN workstation with behavior STANDARD. The misuse example borrows operation from MODEL_INVOKE — that field is excluded from AGENT_RUN.Minimal valid AGENT_RUN workstation:
{
"name": "writer-step",
"worker": "agent-main",
"type": "AGENT_RUN",
"behavior": "STANDARD",
"inputs": [
"ready"
],
"outputs": [
"accepted"
],
"onContinue": [
"ready"
],
"onFailure": [
"failed"
]
}
Incompatible misuse — operation on AGENT_RUN:
{
"name": "writer-step",
"worker": "agent-main",
"type": "AGENT_RUN",
"behavior": "STANDARD",
"operation": "chat.completions",
"inputs": [
"ready"
],
"outputs": [
"accepted"
]
}
operation belongs on MODEL_INVOKE. Keep AGENT_RUN on identity, worker binding, inputs, outputs, and routing fields only.