AI Skills Academy / Course Days 1–3 free

Make an over-privileged AI agent safe to roll out.

Thirty lessons. Each one is a two-minute narrated video and a hands-on lab -- real starter files, a prompt you paste into Claude or ChatGPT, steps to check what it gives you, and an answer key. No cloud account needed. About 21 hours of lab work in total.

The company you'll be working at

Tallyline -- A fictional 45-person B2B invoicing SaaS. Its customers are small accounting firms, so it holds their clients' invoices and bank payout details, and it's SOC 2 audited. Patchbot is Tallyline's AI coding agent. A staff engineer set it up as a weekend project eight months ago and it quietly became load-bearing: it runs in GitHub Actions with an admin AWS key, can run any SQL against production, deploys without CI, and merges its own pull requests.

You've just been hired as Tallyline's first agent platform engineer, with 30 days to make Patchbot safe enough to roll out to every team.

Priya Shah

VP Engineering. Patchbot on every team this quarter, without another incident.

Marcus Lee

Platform / SRE lead. Owns AWS, Terraform, deploys and on-call.

Dana Ortiz

Security engineer (team of one). Owns SOC 2 evidence. Would switch Patchbot off today.

Sam Kim

Staff engineer. Built Patchbot. Defensive, but reasonable.

Already on record

The lint incident: Patchbot fixed a failing lint check by loosening the rule, then merged its own PR.
The migration near-miss: Patchbot ran an ALTER TABLE on production through its SQL tool while debugging. No approval, no migration file.
The Friday deploy: a Patchbot deploy at 17:40 broke invoice PDFs for two hours, and the rollback was a human re-running an old pipeline from memory.

Syllabus

Days 1–10 · The ten disciplines

Day 1: Find where reasoning and authority are collapsedVideo 1:49 · Lab 35 min Day 2: Write (or fix) your agent instructions fileVideo 1:39 · Lab 40 min Day 3: Protect the workflows the agent shouldn't be allowed to touchVideo 1:44 · Lab 35 min

Day 4: Audit what credentials the agent actually inheritsVideo 1:44 · Lab 45 min

Day 5: Replace one broad tool with a narrow oneVideo 1:38 · Lab 45 min

Day 6: List your independent validation gatesVideo 1:35 · Lab 40 min

Day 7: Design the approval screen for a real changeVideo 2:07 · Lab 40 min

Day 8: Reconstruct one execution path end to endVideo 1:49 · Lab 45 min

Day 9: Design the rollback you don't have yetVideo 1:54 · Lab 45 min

Day 10: Score this workflow against the production checklistVideo 1:46 · Lab 40 min

Days 11–24 · The platform

Day 11: Draw your own three-concern splitVideo 1:47 · Lab 35 min

Day 12: Frame the rollout as capability uplift, not replacementVideo 2:02 · Lab 35 min

Day 13: Make your repository legible to an agentVideo 1:50 · Lab 40 min

Day 14: Separate model access from tool authorityVideo 1:52 · Lab 40 min

Day 15: Map your stack onto AgentCore's modular servicesVideo 1:45 · Lab 45 min

Day 16: Check whether your MCP tools still need authorizationVideo 1:43 · Lab 40 min

Day 17: Confirm CI/CD, not the agent, gates promotionVideo 1:50 · Lab 45 min

Day 18: Scale autonomy to consequence, explicitlyVideo 1:59 · Lab 40 min

Day 19: Connect a request to its full traceVideo 1:58 · Lab 45 min

Day 20: Put a number on whether the agent is actually helpingVideo 1:50 · Lab 40 min

Day 21: Check your region's actual authorization boundaryVideo 1:56 · Lab 45 min

Day 22: Write the runbook for your riskiest agent workflowVideo 1:47 · Lab 50 min

Day 23: Set your own baseline before you expandVideo 1:39 · Lab 40 min

Day 24: Build your own service-to-control mapVideo 1:41 · Lab 45 min

Days 25–30 · Capstone: Patchbot v2

Day 25: Build the foundation, before the agent gets write accessVideo 2:06 · Lab 45 min

Day 26: Let the agent look, before it can touch anythingVideo 1:48 · Lab 45 min

Day 27: Let the agent make one small, reviewable changeVideo 1:48 · Lab 40 min

Day 28: Make sure the agent can't pass its own buildVideo 1:35 · Lab 50 min

Day 29: Add the human decision, and the rollback that has to workVideo 1:54 · Lab 50 min

Day 30: Reconstruct the whole run from one correlation IDVideo 1:49 · Lab 50 min

Get the full course

Membership · $49/mo

Days 1–3 are free. Membership unlocks all 30 lessons -- every video, lab, starter file and answer key -- plus live workshops as they're scheduled.

Join — $49/mo

Intro video transcript

Welcome to the thirty day course on governed agentic AI engineering. Over the next thirty days you'll take an AI coding agent that has far too much power, and make it safe enough to roll out to a whole company.

Every lab happens at one company, so the picture builds up instead of resetting each day.

Tallyline is a forty five person invoicing company. It holds customer payout details, and it's audited.

Its agent, Patchbot, runs with an admin cloud key, can run any SQL against production, and merges its own pull requests.

It has already loosened a lint rule to pass a build, altered a production table, and shipped a Friday deploy that nobody could roll back.

You've just been hired to fix that, without taking the agent away.

Each day follows the same shape.

A short video like this one sets up the problem and shows you the evidence.

A full, structured prompt you paste into Claude or ChatGPT. No special tools, no cloud account needed.

And a lab: real starter files from Tallyline, steps to check what the model gives you, an answer key, and a clear finish line.

The thirty days build in three stages.

Days one to ten fix Patchbot itself, one control discipline at a time.

Days eleven to twenty four design the platform around it, from tool authorization to rollout metrics.

And days twenty five to thirty rebuild it from scratch as Patchbot version two, ending with a run an outside auditor can reconstruct.

Two things before you start.

Open a chat assistant in a new tab.

And start one document to keep every artifact you produce. By day thirty, that document is your evidence package. See you on day one.