Image
Workflow
Operations

Runbook execution with a human gate on destructive steps

Runbooks are written down, then run by hand in the middle of the night, because nobody trusts automation with the step that deletes something. An agent runs the safe steps and stops at every destructive one until a named person approves it.

The short answer

Short answer

An agent runs a runbook by working through its written steps in order: it performs the read-only checks and safe actions itself, records each result, and stops at every step that deletes, restarts, rolls back or changes production.

A named person reviews that step and approves or refuses it. The agent never improvises a step the runbook does not contain.

Today

Why are runbooks still run by hand?

The runbook exists. It was written after the last incident, and it lists the checks, the commands and the order. But the person on call still runs it by hand, line by line, because one of the steps restarts a database or clears a queue, and nobody wants automation doing that unsupervised. So the safe steps wait on a tired human too, and the whole procedure takes as long as the slowest person at the keyboard.

Who is this for?

An IT lead, operations engineer or managed services lead whose team keeps written runbooks for recurring operations: clearing disk space, rotating credentials, restarting a stuck service, failing over a replica. It suits teams whose runbooks are specific enough that two people would run them the same way. Writing and changing runbooks, and deciding when to run them, stay with the team.

What changes when the safe steps run themselves?

Part of the runRun by handRun by an agent with gates
Read-only checksTyped one by oneRun in order, with each result recorded
Destructive stepsRun by whoever is at the keyboardHeld for a named person's approval
What the approver seesWhatever they check themselvesThe step, its target and every check result so far
An unexpected resultHandled on instinctThe run stops and pages the on-call person
The recordShell history, if anyone saves itEvery step, result and approval, with names and times

How it works, step by step

The runbook is the program and the gates are where people decide. The agent runs nothing that is not written down.

01

Label each step by risk

With your team, each runbook step is labelled: read-only, safe to perform, or gated. Anything that deletes, restarts, rolls back, changes configuration in production or touches customer data is gated by default.
02

Start the run on a trigger you approve

A run begins when a named person starts it, or when an alert you have tied to that runbook fires. The agent tells the on-call person the run has started.
03

Run the checks and safe steps in order

It performs the read-only checks and safe actions exactly as written, recording each output. This is the repetitive part that Google's Site Reliability Engineering book describes as toil: work tied to running a production service that tends to be manual, repetitive, automatable, tactical, devoid of enduring value, and that scales linearly as a service grows.
04

Stop when anything is unexpected

If a check returns something the runbook does not anticipate, the run stops and the on-call person is paged with the results so far. The agent does not guess at a next step.
05

Human gate: a named person approves each destructive step

Before every gated step, the agent shows the step, its exact target and the check results, and waits. The named approver approves or refuses that step. Approval covers that one step only, so the next gated step waits again.
06

Record the run

Every step, output, approval and refusal is recorded with names and times, and attached to the related ticket or incident.
07

Flag runbooks that need work

Where a run stopped on an unexpected result, the agent lists the step for the runbook's owner to update.

Never in a runbook

What does the agent never do in a runbook?

Improvise

It runs only the steps written in the runbook.

Skip a gate. Gated steps wait for approval every time.

Approve itself

Approval comes from the named person, not from a timeout.

Change the runbook

Owners edit runbooks, and a change takes effect only once approved.

Replace on-call. Being on-call stays with people. The same book calls it a critical duty that many operations and engineering teams must undertake in order to keep their services reliable and available.

Where this fits

Workflow / Automate a Workflow

Running one runbook from trigger to record is a single process: Tier 1, Automate a Workflow. Runbooks specific enough to run the same way twice are what the readiness checklist looks for. A good way to start is shadow mode: the agent runs only the read-only checks beside your engineer and its results are compared before any safe step is handed over. Alerts that trigger runbooks often come from exception detection and routing across your ops stack. More in the use case library.

How do safe steps differ from gated steps?

QuestionSafe or read-only stepGated step
ExamplesCheck disk usage, read logs, list processesDelete files, restart a service, fail over, roll back
Who runs itThe agentThe agent, only after approval
What happens if it is wrongNothing changesSomething important can break
Who labels itYour teamYour team, gated by default when unsure
Tier 1 / Prove
How It Works

Questions, answered

Only if the runbook's owner relabels it, after it has run correctly under approval enough times for the team to be comfortable. The relabel is recorded and can be reversed.
AI Opportunity Audit · Runbooks

Run the safe steps and gate the rest

Bring the runbook your team runs most often and mark which steps scare you.