Cloud Action Agents
How Action Agents run on ActionFlows servers, and how an Action Agent session differs from an ActionFlow run.
Cloud Action Agents
An Action Agent is the hosted conversational agent in ActionFlows. Cloud describes where its runtime executes: on ActionFlows servers rather than in the browser or on a process you host.
Cloud is not a second agent product and does not change the product name. Use Action Agent for the saved conversational configuration and ActionFlow for a saved graph.
The browser is the Action Agent Studio client. Model calls, Knowledge Base search, tools, MCP calls, and session transcript handling run in the ActionFlows cloud runtime.
What cloud execution means
Configure and test an agent in Action Agent Studio, then set it to Active when outside callers and triggers should be able to start new sessions.
After activation, an agent can receive work from:
- a Slack, Discord, Teams, or email trigger;
- a scheduled trigger;
- a backend using REST, the TypeScript SDK, or hosted MCP; or
- an Action Agent node inside an ActionFlow.
The runtime remains server-side. You do not keep a browser tab open or run the model on your laptop.
ActionFlow and Action Agent are different units
| Concern | ActionFlow | Action Agent |
|---|---|---|
| Configuration | A saved graph of nodes, edges, inputs, and selected resources | A saved prompt, model, capabilities, and trigger bindings |
| Entry point | A supported start node, then a Flow run | A new Agent session, a message on an existing session, or a trigger-created session |
| Primary identifier | Run history row and runtime run ID | sessionId for the conversation; each message produces turn activity |
| Work unit | One sequential graph execution, with node results and a terminal or paused state | One or more messages and model/tool turns inside a durable conversation |
| Human interaction | A HITL node can pause the graph for approval | The conversation itself is the human-facing surface; tools and HITL-like capabilities are separate |
| Usage surface | ActionFlow Usage and settled credits for the run | Agent Usage and model/turn usage for the session's turns |
| Lifetime | The run ends or pauses; a HITL continuation is a new admitted run | The transcript can continue across messages and runtime suspension |
A channel event currently creates a new Agent session; it does not reopen a session by thread ID. Store and reuse a sessionId only for a conversation that your integration owns.
Active and passive Agents
Agents use the same product idea as ActionFlows, with a different execution boundary:
- Action Agent Studio chat can start while the Agent is passive so you can test configuration.
- REST, SDK, hosted MCP, channel triggers, scheduled triggers, and an Action Agent node need an active Agent when they start a new session.
- An active Agent still needs a usable model, provider credentials, organization credits when applicable, and ready tool or trigger connections.
Deactivation blocks new external starts and trigger ingress. It does not revoke an existing public sessionId in the current runtime. See the Action Agent lifecycle for the exact state and readiness rules.
An Action Agent inside an ActionFlow
An ActionFlow can contain an Action Agent node. The outer graph is still an ActionFlow run. The node can start an isolated Agent session or continue a persistent session, depending on its saved keepSession setting.
Action Agent node contract
The Actionflow Studio selector lists only active Agents. Configure the Agent, prompt, and session mode in Node settings. The prompt input is required. The node trims the prompt, rejects an empty value, and rejects a value longer than 100,000 characters.
The node dispatches the prompt and does not wait for the Agent's answer. It returns as soon as the prompt is queued on the Agent chat, and the Agent's turn runs on its own durable chat run. The Agent's reply belongs to the Agent session, not to the Flow run.
- With
keepSession: false, the node starts a fresh isolatedapisession for the dispatch and leaves it open so the turn can finish and stay readable in the Agent's session list. - With
keepSession: true, the node finds or creates an openapisession for the same organization, keyed byactionFlowId,nodeId, andagentId, and reuses that conversation for later runs of the node.
Before it dispatches, the runtime checks that the Agent is active again. This differs from public continuation through /v1/chat/completions and hosted MCP: a public session checks active when it starts, but the current runtime does not check active again when an existing public session receives another message. See Sessions & chat.
Each dispatch carries a stable message id built from the run history, node, retry attempt, and loop context, so a retried attempt cannot queue the same prompt twice. Cancelling the Flow run stops the Agent only while the node is still running; once the prompt is dispatched, cancelling the Flow leaves the Agent turn alone.
The node has no outputs. Invalid input, an inactive or missing Agent, session errors, and dispatch failures become node failures. Agent-side failures that happen after the dispatch are not visible to the Flow run; they appear in the Agent's logs and session instead.
Configuration and secrets
Cloud hosting does not create a separate configuration system. The Agent still owns its model, prompt, Knowledge Base collections, skills, tools, MCP servers, and triggers.
Provider keys, organization connections, outbound MCP authentication, and internal runtime credentials stay on the server. Public Agent metadata does not expose the system prompt, tool configuration, MCP configuration, or configVersion. See Credential management.
Credits
Platform-backed Agent turns check organization credits and record usage on the Agent's Usage surface. An Agent using an organization-owned provider key uses BYOK and does not use ActionFlows platform chat credits; external providers and tools may still have their own charges.
Do not infer Agent usage from an ActionFlow estimate or from the existence of a session. Use the Agent Usage and session/turn records. See Usage & cost.
Related
What is ActionFlows AI
An introduction to ActionFlows AI: the platform for building AI-powered automations visually with nodes, triggers, integrations, and models.
Node Anatomy
How current ActionFlow nodes are configured, connected, executed sequentially, and represented in live output and run history.