A procedure that assumes the normal case is where automation gets dangerous
Runbooks describe the normal path, which is what a script already handles. The reason a person runs them is step four, where the output is not what the document expects, and someone has to decide whether to continue.
At that point, a script either fails or continues regardless, and continuing is how a routine procedure becomes an incident. The fault is not in the automation. The fault is in the assumption that the document covered everything.
An agent can hold the intent of a step rather than its literal command, which means it can recognize an unexpected result instead of only a non-zero exit code.
What it must then do is stop. The value is in running the tedious eleven steps and refusing the twelfth, with the state it reached written down for whoever picks it up.
How the agent is built in Agent Studio
Each runbook lives in a prompt skill, with every step stated alongside the command, the expected result, and what a difference would mean.
A second prompt skill declares the boundary: which steps the agent may perform, which always require a person, and which it may never take under any circumstances.
The system prompt requires the agent to stop on any deviation rather than interpret it. The agent must report the state it reached and what it observed at the step where the deviation occurred.
It forbids improvisation entirely. A step not written in the runbook is not performed, even where the agent believes it would help.
An HTTP Request node specifies the permitted actions, and the Agent Chat node maintains a durable session so that a long procedure keeps its place throughout the run.
Every step is posted to Slack as it is completed, so someone watching sees progress rather than a silent gap followed by a result.
A Human in the Loop node gates every step marked as requiring a person. Google Sheets records each run and its deviations, so the runbook itself improves.
What this agent is built from
- Prompt skills: each runbook with the intent of every step and its expected result.
- Prompt skills: which steps are permitted, which need a person, and which are never taken.
- System prompt: stop on deviation, report the state reached, and do not interpret.
- System prompt: no improvisation. An unwritten step is not performed.
- Agent Chat (Util): a durable session that persists throughout a long procedure.
- Slack (Communication): each step is posted as it is completed, so progress is visible.
- When an ActionFlow is enough: If the procedure has no branches and no judgments, build the workflow. It should remain a workflow.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.