Agent nodes

An LLM node makes one call and returns the answer. An agent node is for steps that need to work at it — investigate, call tools, read the results, and decide what to do next, repeating until the job is finished. It’s agentic mode as a single step inside a workflow.

When to use one

Use an agent node when the step can’t be done in one shot:

• “Research this company and produce a one-page profile” — needs several searches and a judgement call on what matters.
• “Reconcile these two reports and flag the discrepancies” — needs to fetch, compare, and decide.

If you already know the exact prompt-and-response shape, a plain LLM node is simpler and cheaper. Reach for an agent when the step needs to find its own path.

Giving it tools

An agent can only use the tools you allow it. You provide an allowlist of tools — web search, code execution, your connected integrations — and the agent calls them as needed. An empty allowlist means the agent has no tools and behaves like a plain model call, so granting tools is a deliberate, opt-in choice.

Caution

Scope agents tightly. Give a node only the tools its task needs — not everything you have connected. It keeps runs predictable and avoids an agent wandering into systems it shouldn’t touch.

Guardrails

Because an agent runs in a loop, it has limits so it always finishes:

• Max iterations — the most tool-call rounds it may take before it has to stop.
• Max time — a wall-clock ceiling on the whole step.
• Stop condition — an optional signal that ends the loop early once the goal is met.

When a run ends, the node reports why it stopped — finished naturally, hit the iteration limit, hit the time limit, was cancelled, or matched the stop condition — so you can tell a clean finish from a truncated one.

Watching it run

As the agent works, each iteration streams into the runner: the tools it called, the arguments it used, the results it got, and how long each took. You can cancel a run at any time and the loop unwinds cleanly. The node’s output includes the final result plus the full list of tool calls it made, so an agent step is auditable like any other.

Agent node vs. a sub-workflow

Both let you build bigger things from smaller pieces. Use an agent when the step needs judgement and an open-ended set of moves; use a sub-workflow when the steps are fixed and you just want to reuse them.