#Give an agent tools

An agent can call flows, adapter actions, MCP servers, its knowledge vaults and its own tasks; each tool reaches the model with a name, a description and the fields it expects.

#Kinds of tool

flowchart LR
  M["The model"] --> F["Flow<br/>a BPMN tool — the agent<br/>pauses until it has run"]
  M --> A["Adapter action<br/>a call to one of your systems"]
  M --> P["MCP server<br/>its tools"]
  M --> K["Knowledge vault<br/>read, and write if allowed"]
  M --> T["Own task<br/>task_&lt;name&gt; — pauses too"]

Kind	Added on	Runs
Flow	Tools tab → Add tool → Model a BPMN flow	As a process of its own; the agent waits for it
Adapter action	Tools → Adapter action — System, Adapter, Actions	Directly, as a call to the system
MCP server	Tools → MCP server	The server's tools
Knowledge vault	Knowledge tab — see knowledge vaults	Read tools; write tools only for a vault set to read & write
Own task	Tasks tab — see its own tasks	As a run of that task; the agent waits for it

#What the model sees

For every tool the model gets a name, a description and the fields it has to fill in:

A flow: the flow's name; the description of its start event — this is what the agent learns about when to use it; and the fields of the start event's input schema. Fields you preset in the flow are left out, and the preset values are used, whatever the model sends.
An adapter action: <adapter>_<action>, the action's description, and its parameters and request-body fields.
An MCP tool: what the server says about it.

Write the descriptions for the model: what the tool does, and when to use it.

⚠️ Two adapter actions whose names come out the same replace each other without a warning.

#Add an MCP server

In the dialog, give a Name, the URL — such as https://mcp.example.com/mcp — and the Transport (HTTP, recommended, or SSE), then how to log in: None, Bearer Token, API Key or Basic Auth with a credential of a system assigned to the agent, or OAuth. Test & select tools connects and lets you choose which tools the agent may use.

Turn an MCP server off with its switch to keep it configured without offering it: the agent does not connect to it and does not see its tools until you turn it on again.

⚠️ An OAuth server is connected without a login when the agent runs.

⚠️ A server that cannot be reached, or refuses the login, is silently left out of the run.

#When a tool fails

The model is told what went wrong and carries on — it can try again, try something else, or say it could not finish:

an adapter action returns the error;
a flow or task that fails returns the error and what it had produced.

When the model calls several flows or tasks at once, only the first runs; the others are reported back as not executed.

#Limits

❌ No approval step: an agent calls its tools without asking anyone.
⚠️ If a process retries a failed agent step, the agent's calls to your systems are made again — give agents actions that can safely be repeated.
An adapter action waits as long as its adapter's timeout, 30 seconds unless set; a whole agent step in a process at most 5 minutes.
A task may restrict which tools it uses (Restrict on the Tasks tab). ⚠️ In a chat every tool is available.