Action Agents
Durable AI chat assistants you design in Studio: triggers, knowledge, model selection, skills, MCP, tools, security, and artifacts.
Action Agents
An Action Agent is a durable AI assistant owned by your organization. You design it in Action Agent Studio, then chat from the Studio dock or from your backend via the REST API and TypeScript SDK.
Open Studio from Action Agents (route: /action-agent-studio/{actionAgentId}). Left nav opens Settings, Sessions, Artifacts, and Usage. Chat is a dock on the right, not a trigger you configure.
Use an ActionFlow when you need a graph of nodes. Use an Action Agent when you need an ongoing conversation with tools.
Studio layout
Main column, top to bottom:
- Triggers: Slack, Microsoft Teams, Discord, Email, Scheduled
- Knowledge: Knowledge Base collections
- Model selection: system prompt, provider or platform model, effort, fast mode
- Skills: organization and platform instruction packs
- MCP: outbound MCP servers the agent can call
- Tools: builtin tools and named HTTP APIs
- Security: PII, moderation, prompt policy
Generated files: Artifacts.
┌────────┬───────────────────────────┬──────────┐
│ Nav │ Triggers │ Chat │
│ set │ Knowledge │ dock │
│ sess │ Model selection │ │
│ art │ Skills / MCP / Tools │ │
│ use │ Security │ │
└────────┴───────────────────────────┴──────────┘Saving
Each successful save increments configVersion. Edits apply on the next message. An in-flight turn keeps the configuration it started with.
Chat
Start a session in Studio or with the Action Agents API. Each user message runs a turn against the model and tools. Replies can be a full string or SSE text deltas. Credits are gated like flow runs.
External apps authenticate with an organization API key. Trigger.dev stays on ActionFlows servers.