Agent: Auto-Logging & Self-Termination

Unlike a plain HTTP call, the Agent node doesn't just make the AI request and move on — it reports back and checks the run's health automatically, every time. This is a big part of why AI steps are accountable "for free," and it's worth knowing what happens so you don't duplicate it.

It logs itself

After each AI call, the Agent node logs the step to the platform on its own. It reports the agent slug, the model and provider, the token usage (input and output), the duration, the status, and the full request and response. You don't add a Log node after an Agent node — that would double-count. The platform takes those token counts and computes the dollar cost from current model pricing, so the agent's spend lands in your execution log and analytics without any effort on your part.

(This is also why the Log node doesn't support the agent worker type — agents log themselves. The Log node is for the other steps you build.)

It self-terminates on a breach

Logging a step isn't one-way. The platform's response to the Agent node's log call carries the run's running tally — time consumed, cost consumed, the limits, and an abort flag. If the run has hit its cost or time limit, abort comes back true, and the Agent node terminates the run right there: it tells the platform to mark the execution terminated, and stops — so the workflow doesn't keep spending past its budget.

That makes the Agent node one of the two enforcement points for the operator's guardrails (the Log node is the other). You don't wire up limit-checking yourself — placing Agent (and Log) nodes is the enforcement. And it's belt-and-braces: if for some reason a node's self-termination doesn't fire, the platform's background reconciler still stops a run that's over its limit. The mechanics of the abort flag itself are covered in Handling the Abort Flag; here the point is simply that the Agent node does this for you, automatically.