How agents take your runs

Some workflows have agent steps: steps that need judgment, reading or writing instead of a plain tool call. Those runs need an agent.

If you’re in a chat, your chat acts as the agent.
If an automation starts a run and no one is chatting, the run waits for one of your agents.
No agent set up → the run goes to Action items and you finish it yourself.
Agent set up → it claims the run and handles the agent steps.

Available agents

You can use any of these. Each runs on your own plan or account (you pay for the tokens).

Agent	Runs on	Needs
Hermes	Your computer (must be on)	ChatGPT plan, Nous Portal or API key
Claude task	Claude’s cloud	Claude plan
ChatGPT task	ChatGPT’s cloud	ChatGPT Plus / Pro / Business / Enterprise / Edu

Set one up in Get started > Set up an agent, or in Settings > Agents.

Assignment rules

Agents only take your runs: runs from your automations, and runs you start with Run now.
First agent to claim a run gets it. One run → one agent.
One run at a time per agent. Runs of one workflow go to agents one after another.
If an agent asks you a question, your answer sends the run back to it.
If an agent can’t finish a run, it won’t be offered to any agent again. It shows up in Action items under Runs the agent could not finish and on the workflow’s Gave up tab.

Repairing failed runs

In a workflow’s Settings tab, press Let an agent repair failed runs.

The agent takes each failed run once, tries to finish it, and suggests a workflow change on the Suggestions tab. Anything it still can’t finish comes back to you.

Where to see agent activity

Run page → shows which agent holds the run. Take it back stops the agent so you can finish it.
Action items → questions the agent asked, runs it gave up on, and Agents stopped taking runs after repeated failures.
Workflow Suggestions tab → changes the agent recommends.
Settings > Agents → all your agents, last check-in (Hermes) and last run taken (Claude and ChatGPT tasks).
To pause agents or set limits → Pause, limits and stop.