Decision Model Agent Tool Gate

Decision Model Agent Tool Gate

"Is GitHub Actions down right now?" needs fresh service data. "What is GitHub Actions?" usually does not. This template gives an agent a Decision Model as a callable tool so it can make that distinction before reaching for an HTTP tool.

What this workflow does

1. UserQuestion supplies a question about GitHub or GitHub Actions.
2. StatusAssistant calls the Check Lookup Needed Decision node through its tool connection, passing the full question in the agent-provided state field.
3. The Decision Model answers the Noul question needs_live_lookup with a yes probability.
4. At 0.8 or above, the agent calls Fetch GitHub Status, a fixed HTTP tool for GitHub's public status summary. Below 0.8, it skips the HTTP call.
5. The agent reports the yes/no decision, its probability, whether it fetched live data, and the answer.

Both the Decision and HTTP nodes connect to the agent's bottom tool handle. They are not steps in the main execution path. The agent invokes them during its tool loop and receives their outputs as tool results.

Setup

Select a tool-capable LLM credential on StatusAssistant and a Decision Model credential on Check Lookup Needed. The HTTP status endpoint needs no credential. The agent can supply the Decision node's state, but the question, rubric, model, and credential remain fixed. The HTTP tool has no agent-provided fields, so the agent cannot change its URL.

Try yes and no

• Yes: "Is GitHub Actions experiencing an outage right now?"
• No: "Explain the difference between Git and GitHub."
• No: "Write a short introduction to GitHub Actions workflows."

Inspect the Agent trace to see check_lookup_needed followed by fetch_github_status for the yes path. The no path should contain the decision call without an HTTP call. Exact probabilities vary by model.

Handling uncertainty

If the tool errors or returns no probability, the agent reports that the check failed and does not claim current service status. Below the threshold, a question that still needs live information gets an explicit "not verified" response instead of an invented current status.

This is an agent-directed tool selection example. The yes/no threshold is enforced by the agent's instructions, not by a hard backend permission boundary. For consequential actions, put a Condition node in the main workflow before the action. Here the optional action only reads a public service-status endpoint.