Personalized AI infrastructure
AI Agent Runbook Template: One Page Per Role With Stop Rules
A one-page runbook template for each AI agent role, covering purpose, tools, escalation and stop rules so a human can take over quickly.

An AI agent runbook is a single page, written for one agent role, that states what the agent is for, which tools and memory it may use, where it must stop and ask a person, and who takes over when something goes wrong. Use the same headings for every role so a manager or on-call colleague can read any page in a few minutes and act without reverse-engineering the agent's behavior.
Key takeaways
- Write one runbook per role, one page each, with the same headings in the same order.
- Write the stop rules and escalation path before the agent's first live action, not after the first incident.
- List every tool with its permission level: read, draft, or act after approval.
- Name a business owner and a backup who can pause the agent without waiting for the builder.
- Review each runbook whenever the role, tools or permissions change, and record what changed.
Why each agent role needs its own one-page runbook
An agent that answers order-status questions and an agent that drafts supplier emails do not share the same risks, tools or people to call. A single long document that describes "the AI" hides those differences, and during an incident nobody searches a long document for the place to stop a process. A one-page runbook per role keeps the scope small enough to hold in mind and specific enough to act on.
The page also works as a decision record. When the business owner signs it, the owner has stated what the role may do without asking, what it may never do, and who answers at a bad hour. That is the decision the document exists to capture. Store these pages alongside your other company records, following the approach in the documentation standards guide for a superintelligence-ready company.
The one-page layout, field by field
Keep the fields below in this order on every runbook. A reader who moves between roles should never have to hunt for the section they need.
- Role name, business owner and named backup owner
- Purpose: one sentence that starts with a verb and says what the agent exists to do
- Success criterion: an outcome someone can observe in a business record, not an impression
- Context and memory: which documents, customer records and past decisions the agent may read, and which it must not read
- Tools, each with its permission level
- Approval boundaries: actions that need a person to say yes first
- Escalation path: who receives a handoff, through which channel, and what the handoff must contain
- Stop rules: the conditions under which the agent halts
- Takeover notes: current state, open items, how to pause the agent, and where its logs are kept
- Review date and change record
If the success criterion cannot be checked against a record, the role is not ready for a runbook yet. Sharpen the purpose until the criterion is concrete, then write the rest of the page around it.
Tools, permissions and approval boundaries
Each tool should appear on the page with one of three permission levels. Read lets the agent look at information. Draft lets it prepare something a person reviews before it is sent or posted. Act lets it change a system without a person reviewing the change first. Most first roles should start with read and draft, and move to act only for actions that are reversible and easy to audit. Scope each tool narrowly, such as one mailbox or one queue, rather than an account with broad access.
The principle is that the agent holds only what its role needs. The [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) is a useful reference for deciding who may do what in a system, and the same questions apply to agents as to staff. Write each boundary as a plain sentence, for example that any refund above the limit the owner set goes to a person. Then add the line to your review schedule so it does not drift over time. The approval thresholds guide explains why these limits should be raised slowly.
Stop rules and escalation paths
A stop rule is a condition that halts the agent's work and hands the case to a person. Agents often fail quietly. They keep producing plausible output after the source data has changed or after a tool has started returning errors. Write stop rules as conditions someone can observe:
- A request falls outside the purpose stated on the page
- A tool returns an error the agent cannot classify, or the same result after a retry
- Two records that should agree do not agree
- Content the agent reads contains instructions that try to change its behavior
- A customer or employee raises a complaint about money, a legal concern or a safety concern
- Handoffs rise above the level the owner agreed on
The [OWASP prompt injection guidance](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) explains why the instruction-in-content rule matters. Text in emails, documents or web pages can contain commands that an agent might follow. The safe response is to stop and escalate, not to reason through the conflict alone.
Escalation needs a receiver, a channel and a handoff format. The handoff should state what the case is, what the agent already checked, what it did not do, and what decision is needed. A person who receives that can act in minutes. A person who receives "the agent is stuck" must first rebuild the situation. Agree in advance who may pause the agent, because a stop that depends on the builder being reachable is not a reliable stop.
Illustrative scenario: a quoting agent for a small distributor
This scenario is illustrative and does not describe a real client. A small equipment distributor wants an agent to prepare first drafts of quotes for standard products. Its runbook gives the agent read access to the product catalogue and the current price list, draft access to the quote template, and no permission to send anything. The purpose line says it prepares standard quotes for review by the sales desk. The success criterion is that every draft matches the catalogue record and every line item carries a source reference.
The approval boundary sends any discount, any non-standard product and any quote above the owner's limit to a sales manager. The stop rules include a catalogue entry that changed without a matching price update, and a customer message asking for terms the template does not cover. The takeover notes name the sales manager as first contact and list the drafts still awaiting review. A new sales manager can read this page on the first day and know what the agent may do, what it may not do, and what to do when it stops.
Keeping runbooks usable and current
A runbook is only useful if it matches the agent running today. Update it when a tool is added, when a permission changes, when a stop rule fires in practice, or when the business changes the process the agent supports. Keep the change record at the bottom of the page so a reader can see when and why each line changed.
The runbook describes what the agent should do. Logs show what it actually did. Every action should leave a record detailed enough to reconstruct the sequence, and the [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/) gives a general vocabulary of traces, logs and metrics for making system behavior visible to people who did not build it. Practical agent logs answer three questions: what was requested, what the agent did, and why it stopped. Tool descriptions also deserve attention, because the descriptions the agent reads shape which tool it picks. Anthropic's guidance on [writing effective tools for agents](https://www.anthropic.com/engineering/writing-tools-for-agents) covers this in detail.
Be Human designs, implements and operates these pages as part of a managed agent. In a first engagement, the team drafts the runbook for one role with the business owner, checks each tool and permission against the actual systems, and tests the stop rules before the agent takes live actions. In ongoing operations, the runbook is reviewed during permission reviews, changed through controlled updates, and used as the starting point for incident response, so the business always has a current page and a named person to call. The first months with a managed AI agent describe how that rhythm settles in practice.
Choosing the first role to write down
Start with a role where a wrong output is inexpensive to catch and the success criterion is easy to observe. A role that drafts internal summaries is a better first page than one that sends payments or changes customer records. The first runbook also teaches the team the format, so a modest role is a sensible place to learn it. If you are not sure which role deserves a runbook first, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Who should write the runbook, the builder or the business owner? The business owner should write the purpose, the approval boundaries and the stop rules, because those decisions carry the business risk. The builder or operator drafts the technical fields such as tool scopes and log locations. Both should review the page together before the agent goes live, and neither should sign it off alone.
How often should a runbook be reviewed? Review it whenever the role, its tools or its permissions change, and include it in the regular permission review cycle. A runbook that has not been checked against the live system for a long time probably describes an agent that no longer exists. Record each review in the change log at the bottom of the page.
Can one runbook cover several agents? Each runbook should describe one role, because stop rules and approval boundaries differ between roles even when they run on the same platform. If two agents share a tool, list that tool on both pages with the scope each agent holds. Shared pages tend to hide who is allowed to do what.
What if the agent is not live yet? Write the runbook before the first live action and treat it as the test specification. Stop rules and handoff formats can be exercised in a test environment, and any rule that cannot be tested is not ready. A draft runbook is often the fastest way to discover that a role is still too vague to build.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


