Working with other agents
How an Action Agent finds and messages other agents in the same organization, updates its own studio settings, and lists, reads, runs, or creates ActionFlows.
Working with other agents
Every Action Agent turn includes organization tools. They are not an MCP server you attach, and the Agents card in Action Agent Studio stays on. The agent can use them when the task needs another agent or an ActionFlow.
These tools stay inside the agent's organization. Template agents and template flows are not included. Another organization's agents and flows are not visible.
Find and message another agent
The agent can list other Action Agents in the organization, read one, and send it a message.
A peer read includes the name, description, active state, system prompt, skill names, knowledge collection names, MCP connection names, trigger channels, and HTTP tool names and methods. It omits credentials, headers, and tool URLs.
A message starts a new session on the target agent and waits for that turn. The target must be active. A message to itself is refused. A message to an agent already in the chain is refused.
A user-started agent can hand work to another agent, and that agent can hand it on once more. The next agent cannot start another agent, and it cannot run, create, or update an ActionFlow. Listing and reading still work.
Update its own studio
The agent can read and change its own Action Agent Studio settings:
- profile, appearance, and the system prompt the owner wrote
- the primary model, and the routing list
- triggers, preset, skills, MCP connections, database grants, HTTP tools, and the builtin tool
- the lead reviewer and the guardrail policy
- a knowledge collection, including a new collection with a note, attached to this agent
Omitted fields stay as they are. A list, such as skills, tools, triggers, databases, or routes, replaces that whole set. The primary model is still one mode at a time: platform, own-key, or custom.
Reading itself returns HTTP URL templates and still omits stored authentication and headers.
ActionFlows
The agent can list up to 50 ActionFlows in the organization, read one, read a run's status, run an active flow, and create or update a flow.
A flow read redacts input values whose names look like a credential, secret, token, password, API key, authorization value, or header. A run starts only when the flow is active. A flow the agent creates starts inactive, and the graph is checked before it is saved. Edges without nodes are refused.
Run status does not include the run data or trace.
Creating or running a flow spends the same organization credits and follows the same validation as a run started from the API. The agent does not bypass an inactive flow or a plan limit.
Platform instruction
ActionFlows appends a platform note after the owner's system prompt and skills. The note is not written into the Action Agent Studio prompt, so a self read still returns the text the owner wrote.
The default note tells the agent to use an ActionFlow in this organization when automation fits the task, and not to mention the note. An operator with Action Agent update permission can replace that note in Admin → Agent instructions. A blank saved note falls back to the default. The next turn picks up a saved edit.