Business Agent Launch Pack
Build one narrow reusable agent with repository guidance, a custom skill, explicit permissions and human gates, real and held-back evals, a passing run record, and exact launch, update, disable, and rollback instructions.
Use this printable workbook with the free Build a Business AI Agent with AGENTS.md, Skills and Evals course. Complete one section after each lesson, preserve source evidence, and record manual, blocked, repair, narrower-scope, or stop decisions honestly.
What it contains
- Narrow job and binary outcomes
- Appropriate native surface
- Durable repository guidance
- Validated custom skill
- Permission and human-gate matrix
- Real and held-back evals
- Passing run evidence
- Installed-copy proof
- Launch and rollback instructions
How to use it
Use permitted real or sanitized evidence where available. Keep fictional fixtures clearly labelled. Write expected outcomes before tests, leave missing values unknown, and name the person who owns every consequential decision.
The workbook does not establish production readiness, legal compliance, security approval, provider availability, or business results. Those require current evidence and separate approval.
Give the agent one bounded job
The ten-page PDF helps turn a broad idea into a reviewable task. Choose one input, one intended output and a named owner. List the sources the agent may use and the actions it may take. Reading a record, drafting a response and sending it are separate permissions.
For a fictional first exercise, ask for a draft internal brief from an approved folder. Require source links and an unresolved-questions section. Keep the first version read-only so a reviewer can assess its usefulness before granting write access or adding external recipients.
Define a useful failure
Try a missing source, conflicting instructions and an input outside the agreed task. Write what the agent should return, who receives the exception and when it should stop. A plausible answer without evidence may be less useful than a clear request for missing information.
Review representative outputs with the person who owns the work. Preserve examples of acceptable results and the corrections that matter. Use the business AI agent course to carry those decisions into the lessons. The PDF is a design companion; a completed worksheet alone does not establish that an agent is reliable or ready to act on customer records.