KNIME Documentation

Outdated version

This is the documentation for KNIME Analytics Platform version 5.11 - go to the latest version.

Prompt an agent ​

Start and control an agent’s reasoning process using the Agent Prompter or Agent Chat View.

Prompting an agent means providing a task and letting the agent decide how to approach it, including whether to call tools and how to combine results.

When this is useful ​

Prompt an agent when you want to:

  • Let a model reason step by step to solve a task
  • Allow the agent to decide which tools to use
  • Execute multi-step logic within a workflow
  • Build assistants that go beyond single prompts

If you only need one-shot prompting without reasoning or tool usage, see Prompt a model instead.

Choose the right entry point

Use Agent Prompter for workflow-driven, automated, or debuggable execution. Use Agent Chat View to expose an interactive chat interface to users.

How agent prompting works ​

When an agent is prompted, it:

  1. Receives the system message (behavior and rules)
  2. Receives the user message (task or request)
  3. Reasons about the task
  4. Optionally calls tools
  5. Uses tool results to continue reasoning
  6. Produces a response

This reasoning loop is explicit and inspectable.

Prompt an agent step by step ​

Prerequisites

Define the system message ​

The System Message defines the agent’s role, rules, and constraints.

Use it to specify:

  • what the agent is responsible for
  • how it should behave
  • what it should or should not do

Example:

You are a support assistant.
Always verify information using your available tools.
Do not guess when information is missing.

The system message strongly influences how the agent reasons and how reliably it uses tools.

Be explicit and restrictive

Clear rules lead to more predictable agent behavior. Vague system messages result in inconsistent reasoning.

Define the user message ​

The User Message describes the task the agent should perform.

Example:

Check whether a table for two people is available for tomorrow evening.

Each execution of the agent starts from the provided user message.

Start agent reasoning with Agent Prompter ​

Use the Agent Prompter node to trigger the agent’s reasoning loop from within a workflow.

This approach is best suited for:

  • automated execution
  • scheduled workflows
  • debugging and inspection
  • non-interactive agent runs

When executed, the Agent Prompter produces a conversation that includes user messages, agent reasoning, and tool interactions.

Prompt an agent interactively with Agent Chat View ​

Use the Agent Chat View node to allow users to interact with the agent through a chat interface.

With Agent Chat View, users can:

  • type requests or questions
  • receive responses in real time
  • optionally see tool calls and results

This is the preferred option for:

  • assistant-style applications
  • demos and user-facing tools
  • interactive exploration

You can control whether users see internal tool usage or only the agent’s final responses.

Result ​

After prompting, the agent:

  • reasons about the task
  • optionally calls tools
  • produces a response based on its reasoning

The output is a conversation that can be:

  • inspected for debugging
  • stored or logged
  • passed to downstream workflows
  • displayed to users

Important notes ​

  • Agents do not retain memory unless you explicitly manage it
  • Tool availability depends entirely on what you register
  • System messages should be treated as part of the workflow logic
  • Agent behavior is reproducible when configuration is fixed

Next steps ​

Once you can prompt an agent, you can: