Build a Business AI Agent with AGENTS.md, Skills and Evals

Define One Agent Job and Outcome

Choose one repeatable business job, define the useful output, and turn quality into binary evidence.

• Define One Agent Job and Outcome (16 min)

Choose the Smallest Native Surface

Choose among a prompt, AGENTS.md, skill, specialist agent, connector, plugin, or schedule.

• Choose the Smallest Native Surface (15 min)

Write Durable Repository Guidance

Write concise AGENTS.md guidance that identifies context, sources of truth, boundaries, and conditional detail.

• Write Durable Repository Guidance (17 min)

Build a Concise Custom Skill

Create a discriminating skill with progressive detail, deterministic helpers, and trigger tests.

• Build a Concise Custom Skill (20 min)

Set Permissions and Human Gates

Define reads, writes, external actions, secrets, limits, approvals, and human decisions.

• Set Permissions and Human Gates (18 min)

Create Real and Held-Back Evals

Build structural, deterministic, behavioural, boundary, hostile-input, and discovery tests.

• Create Real and Held-Back Evals (20 min)

Run, Grade, Repair, and Record

Execute the same cases, grade independent outcomes, diagnose failures, and preserve run evidence.

• Run, Grade, Repair, and Record (18 min)

Verify Launch, Rollback, and Maintenance

Test the installed copy and finish invocation, update, disable, rollback, and maintenance instructions.

• Verify Launch, Rollback, and Maintenance (18 min)

What you will work towards

Work through this course to prepare a launch pack for one bounded business agent. The curriculum takes you through durable repository guidance, a concise skill, permission boundaries, evals and operating instructions.

Start with one recurring job, a working folder and an example of an acceptable output. Use synthetic information while learning when permission to use live records is unresolved.

A practical case to carry through the lessons

An agent preparing a weekly account brief needs approved source files and a clear output format. It does not need permission to contact the customer or change the CRM.

Hold some evaluation cases back while building. A repair that passes the example used to write the instructions still needs a fresh check.

How to know the work is ready for review

Before calling the exercise complete, check that another operator can run, grade, disable and roll back the agent using the recorded instructions.