Direct model calls

An InferenceRequest is a bounded, sealed model call outside any conversation: no history, no tools, no turn. Use it for classification, extraction, embeddings, speech and image work, or any model operation where your application owns the whole request.

POST /agent_api/v1/workspaces/{workspace_id}/inference_requests
GET  /agent_api/v1/workspaces/{workspace_id}/inference_requests
GET  /agent_api/v1/workspaces/{workspace_id}/inference_requests/{id}
POST /agent_api/v1/workspaces/{workspace_id}/inference_requests/{id}/cancellation
GET  /agent_api/v1/workspaces/{workspace_id}/inference_requests/{id}/events
POST /agent_api/v1/workspaces/{workspace_id}/inference_requests/input_estimate

Workloads

Discover usable models per workload with GET /agent_api/v1/models?workload=…:

Workload	Purpose
text_generation	Text replies, with optional reasoning controls and image or PDF inputs on capable models
image_generation	Generate or edit images
speech_generation	Text to speech
transcription	Speech to text
embedding	Vector embeddings

Each model's definition declares its capabilities, limits and optional pricing; a workload lists only models whose provider is enabled, credentialed and visible.

Lifecycle

1. Create with the model reference, the request body for that workload and an Idempotency-Key. The request is sealed as submitted; later model changes never rewrite it. Creation has its own documented replay response.
2. Follow execution state over HTTP or the request's events feed; the events replay with the same cursor and watermark contract as conversations.
3. Read the result and recorded usage. A classifier refusal or exhausted retryable transport/server failure may advance through the captured ordered fallback list, within the request's model/provider constraints and finite attempt budget. The request reports the model that actually answered; see model recovery.
4. Cancel through …/cancellation when the result is no longer needed.

input_estimate is a write-free advisory that counts the request's input against the model's limits before you spend on it.

Gateway behavior

The same catalog, credential, reasoning, prompt-format adaptation and pricing rules apply as for conversation replies: the request compiles against the model's definition, reasoning uses reasoning_enabled and reasoning_effort, and usage is recorded with known cost in the account's unit. Rate limits are per resource (6,000 requests per minute by default, shared by create, list and read).

When to use a conversation instead

If the call should remember earlier exchanges, use tools, ask a person, run in the background or report back later, it belongs in a conversation turn or a standalone run. See Orchestration.

Request and response bodies for each workload, the replay contract and status codes are in the InferenceRequests reference; model discovery and reasoning selection are in Models and providers.