API

Agents API

An agent is the unit of work: a system prompt, a model config, an optional tool allow-list, and optional structured JSON output. Every edit is versioned, so you can inspect history and roll back.

Endpoints

POST   /v1/agents                              create
GET    /v1/agents                              list
GET    /v1/agents/{id}                         fetch one
PATCH  /v1/agents/{id}                         update (creates a new version)
DELETE /v1/agents/{id}                         delete
POST   /v1/agents/{id}/activate                status -> ACTIVE
POST   /v1/agents/{id}/deactivate              status -> INACTIVE
POST   /v1/agents/{id}/execute                 run it
GET    /v1/agents/{id}/versions                version history
GET    /v1/agents/{id}/versions/{version}      one version
POST   /v1/agents/{id}/rollback                roll back to a version
GET    /v1/agents/stats/summary                counts + run stats

Configuration

The config object holds the system prompt and model settings. Lower the temperature for classification and triage; raise it for drafting and content.

{
  "system_prompt": "You are a lead-qualification assistant. Score each lead hot, warm, or cold and suggest the next step.",
  "model_config": {
    "provider": "anthropic",
    "model": "claude-sonnet-4-20250514",
    "temperature": 0.3
  },
  "max_tokens": 800
}

Running an agent

POST /v1/agents/{id}/execute
{ "inputData": { "message": "Where is my order #12345?" } }

Input resolution follows one rule: if the agent has an input_fields allow-list, only those keys are used; otherwise the first message/prompt-style field is used verbatim; otherwise the whole input object is passed as JSON.

Agents start INACTIVE. An inactive agent will not run — activate it first.