Action types

Eighteen action types cover what enterprise agents actually do. The enum is closed: an unrecognized type fails schema validation, so pick the closest fit rather than inventing one. unknown exists for genuinely unclassifiable actions.

Data actions

Type Use for
read Reading a record, file or resource without changing it
create Creating a new entity
update Modifying an existing entity
delete Deleting an entity
write Writes where create vs update is ambiguous

Communication actions

Type Use for
send Sending a message or document to an external recipient
download Pulling data out of a system
upload Pushing data into a system

Execution actions

Type Use for
execute Running a workflow, job or command against a target system
call_model A call to an LLM (chat, completion, embedding)
call_tool Invoking a tool, plugin or function
call_agent Delegating to another agent

Control actions

Type Use for
login Authenticating or establishing a session
grant Granting access, role or permission
revoke Revoking access, role or permission
approve Human-in-the-loop or policy approval
escalate Handing off to a human or higher authority

Choosing well

The action type is what findings and filters group on, so precision matters at the category level even when the target is coarse:

  • A copilot reading a customer record is read with target_system: crm, not execute.
  • An agent emailing a customer is send with external_recipient, not execute with operation email.
  • A multi-step RPA job against the ERP is execute; the individual record changes inside it are better as their own create/update events if you can observe them.
  • Model and tool calls get their own types (call_model, call_tool) because token usage, latency and failure classes attach to them and power the quality analysis.

If your agent framework emits OTel GenAI spans, the SDK maps them for you: execute_tool becomes call_tool, invoke_agent becomes call_agent, chat/completion/embeddings operations become call_model.