Agent (AGENT)
An AI model that answers the incoming message. It can use the tools wired to its Tools port. It is the only module with a Tools port, so it is the centre of every pipeline that uses tools.
Category: Models & Agents · Deployed: yes · Tool-capable: no (it uses tools)
Ports
| Port | Kind | Purpose |
|---|---|---|
Prompt ( |
Input |
The message or task |
Tools ( |
Input |
Wire tool modules here. Each one becomes a tool the model can call |
Response ( |
Output |
The model’s answer |
Properties
| Field | Required | Meaning |
|---|---|---|
Description ( |
No |
A short note on what this agent is for. The model also reads it as part of its role. Put the real rules into Instruction |
Instruction ( |
No |
The agent’s job and rules: what to answer, which tool to use when, in which language. This text has the biggest effect on the agent. More precise text gives more predictable answers |
Model Provider ( |
Yes |
Exactly one of |
Model Name ( |
Yes |
The model, for example |
Model URL ( |
Yes |
The address of the model server, for example |
API Key ( |
No |
Needed for |
Model Timeout (min) ( |
No |
How long to wait for one answer of the model, 1–120 minutes. Pre-filled with 20. Higher: slow models and long answers can finish. Lower: a stuck call fails sooner |
Select AI Server & Model
The button Select AI Server & Model fills the model fields from a registered AI server. Choose the server, then one of its models. Apply writes:
-
the provider (
OLLAMA,OPEN_AIorGEMINI; an Anthropic server is entered asOPEN_AI), -
the model name,
-
the server’s address, and
-
the server’s API key, if it has one.
This is a copy. If you change the AI server later, the module keeps the old values. Use the button again to update them.
Model address
-
localhostand127.0.0.1are replaced byhost.docker.internalwhen the pipeline is sent to the agent. With Docker, this name points to the machine that runs the agent’s container. With Kubernetes, use an address the agent can reach inside the cluster. -
OPEN_AI:/v1is added to the address when it is missing. -
ANTHROPIC: the field must be filled, but the address is not used. The model is always reached at Anthropic’s own address. -
GEMINI: the address is used as entered. Enter Google’s address or the address of your proxy.
What happens at run time
-
The agent sees the whole chat session. That includes earlier messages of the chat and the results of the modules that ran before it in this run. Inside an Iterator or a Loop it also sees the earlier items and rounds.
-
Tools: the model decides when to call a tool, and with which values. If a tool fails, the model gets the error message and can react to it. The run goes on.
-
Loop protection: one run may call models at most 500 times. Then the run stops with an error. This stops an agent that keeps calling tools without end.
When the pipeline is checked
Model Provider, Model Name and Model URL must be filled. If one is empty, the app lists it and refuses to
assign the pipeline to an agent. Other mistakes show up only when the pipeline is sent to the agent, for
example a provider in small letters, or GEMINI without an API key. The agent then refuses the pipeline
with an error message.
Use cases
-
The answering step of a chat pipeline.
-
A worker with tools: wire API Call, SQL Database, Web Search or AI Topic into its Tools port and let it choose.
-
Several agents in one pipeline, each with its own model: a small, fast one to sort requests, a strong one for the final answer.
-
Asking the user something: the agent’s answer is the question, and the user’s next message is the reply. The chat history carries it. No extra module is needed.