Top

Agentic Automation — Direct Work That Continues Itself

In 2026, AI agents do not just answer questions. They pick up a task, work on it step by step, and come back with a result, such as a merge request on your app. That makes the Product Manager the agent product owner: you define the tasks, set the limits, review the outcomes, and own the backlog agents work from. This course is part of Orchestration in the PILOT Program™, and it is the most agentic one of all.

What you’ll be able to do

Explain the agent loop

Describe plan, act, observe in plain English, and say where an agent can go wrong at each step.

Understand tools and MCP

Know what function calling and the Model Context Protocol (MCP) let an agent touch, so you can scope what it is allowed to do.

Write agent-ready issues

Turn a vague wish into an issue with clear acceptance criteria and a definition of done that an agent can pick up.

Set checkpoints and guardrails

Decide where a human must approve, and which permissions an agent should never have.

Review outcomes, not code

Use a checklist to judge an agent-made merge request by what it does for users, with no code access needed.

Weigh cost, speed and quality

Compare agent runs by cost, latency and evals so you can defend your backlog choices.

Your certification path ($0)

Work through the steps in order. Steps 1 to 3 are the no-code track and give you the whole picture. Steps 4 and 5 are the low-code track if you want to build workflows or understand MCP more deeply. Every step below can be completed for $0. We checked each page before listing it.

Step 1: Hugging Face AI Agents Course (Fundamentals)

Provider: Hugging Face (huggingface.co/learn)
You earn: Fundamentals certificate (complete Unit 1). A fuller Certificate of Completion needs Unit 1, a use-case assignment and the final challenge.
Cost: Free certificate  |  Time: about 8 hours

Track: No-code friendly for Unit 1 (concepts: what an agent is, tools, the think-act-observe cycle). The later units use Python, so read along as a PM or pair with a coach.

Start the course

Step 2: Zapier AI Builder Path

Provider: Zapier Learn
You earn: Certificate of completion (you can add it to LinkedIn)
Cost: Free certificate (free Zapier account needed)  |  Time: about 2 hours

Track: No-code. About 105 minutes. Builds from a first automation toward agentic AI workflows. Zapier also offers a short MCP Path (55 min) at learn.zapier.com.

Open the AI Builder Path

Step 3: Anthropic Academy: Introduction to subagents

Provider: Anthropic (anthropic.skilljar.com)
You earn: Completion certificate issued through Anthropic Academy (confirm on the course page when you register)
Cost: Free course  |  Time: about 1 hours

Track: No-code concepts. A short course (4 lessons) on creating and designing sub-agents, which is the heart of orchestration.

Register free

Step 4: n8n Essentials: Your First Workflows (N8N101)

Provider: n8n Learn (learn.n8n.io)
You earn: Badge and certificate of completion. n8n says this is not a formal certification; a certification program is planned.
Cost: Free certificate of completion (n8n free trial or self-hosting needed for exercises)  |  Time: about 4 hours

Track: Low-code. Visual workflow building. Continue with N8N103 “In Practice: AI, Testing & Best Practices” from the n8n catalog if you enjoy it.

Open N8N101

Step 5: Anthropic Academy: Introduction to Model Context Protocol

Provider: Anthropic (anthropic.skilljar.com)
You earn: Course with a final assessment on MCP. The course page does not describe a certificate, so treat this as a free knowledge step.
Cost: Free course (certificate not confirmed)  |  Time: about 3 hours

Track: Low-code / stretch. It builds MCP servers in Python, so watch the concept lessons (what MCP tools, resources and prompts are) and skip the coding labs.

Register free

Later, optional and NOT part of the $0 path

DeepLearning.AI “Agentic AI” (Andrew Ng, about 10 hours). You can audit it free, but certificates need a DeepLearning.AI Pro membership, listed at $25/month billed annually or $30/month billed monthly. The short courses on crewAI multi-agent systems and MCP with Anthropic are free to take for now, but their accomplishments also need Pro.

See the course

Practice in GitLab

Do these on your real project board, with no code and no new projects. Take your time and share your results with your coach.

Exercise 1: Create the agent:: label set

In your project, open Manage > Labels > New label and add the labels below. Names with a double colon are called scoped labels: an issue can hold only one label from the same scope, so it cannot be “ready” and “needs-review” at once. If you cannot create labels on the shared project, ask your coach, and do not create a new project. See the GitLab docs on scoped labels.

agent::ready          Issue is agent-ready: clear criteria, definition of done, guardrails. Colour: green
agent::in-progress    An agent (or its run) is working on it. Colour: blue
agent::needs-review   Agent has stopped and handed back. Waiting for the PM. Colour: orange
(optional)
agent::blocked        Agent hit a guardrail or question and needs a human answer. Colour: red

Exercise 2: Write 3 agent-ready issues

Create these three issues on your real project, adapting the details to your own product. An issue is agent-ready when someone who knows nothing about your project could finish it from the issue alone: the why, the exact change, numbered acceptance criteria, a definition of done, what is out of scope, and guardrails. Start with these templates, then save the best one as an issue template under Manage > Issue templates (or ask your coach to add it). See description templates.

Issue A – a small feature

Title: Add a "Contact support" link to the footer of every page

Why (user value): Customers cannot find help, and 12 support emails last month asked "where is the help page?".

What to change: Add a link labelled "Contact support" to the footer, pointing to our existing /help page. Do not change the header.

Acceptance criteria:
1. A "Contact support" link appears in the footer on the home, pricing and account pages.
2. Clicking it opens /help in the same tab.
3. On a phone-width screen the link is visible without sideways scrolling.
4. No other text or links in the footer change.

Out of scope: redesigning the footer, writing new help articles.

Definition of done:
- Merge request is open, linked to this issue, and the pipeline is green.
- The review app link is in the MR description with a screenshot on desktop and phone.
- All four acceptance criteria are ticked in the MR.

Agent guardrails: only touch footer files. Do not edit settings, secrets or pipeline files. Stop and comment if /help does not exist.
Labels: agent::ready, type::feature

Issue B – a bug fix

Title: Fix: Password reset email shows the wrong company name

Why (user value): New users think the email is spam and do not click it.

Steps to reproduce: 1) Open the sign-in page. 2) Choose "Forgot password". 3) Read the email subject and greeting.

Expected: Subject "Reset your Ignite Curiosity password" and greeting "Hi <first name>,".
Actual: Subject and greeting use an old product name.

Acceptance criteria:
1. Subject and greeting match the Expected text exactly.
2. A user with no first name gets the greeting "Hi there,".
3. The email still contains a working reset link that expires as before.
4. A test in the pipeline covers items 1 and 2.

Definition of done:
- Merge request linked, pipeline green, a screenshot of a test email attached.
- The agent lists what it changed in plain English in the MR description.

Agent guardrails: do not change how reset links are made or how long they last. Ask a human before touching anything about security.
Labels: agent::ready, type::bug

Issue C – a research task (no code at all)

Title: Research: Compare three options for a weekly progress email (no code changes)

Why (user value): We want learners to get a weekly progress summary, and need a recommendation before building.

Deliverable: One comment on this issue (about 1 page) comparing three options for sending the email.

Acceptance criteria:
1. The comment compares exactly three options in a table: cost per month, effort (low/medium/high), risks, and privacy impact.
2. Every cost figure has a source link, and the date it was checked.
3. The comment ends with one clear recommendation and two open questions for the PM.
4. Nothing is changed in the repository.

Definition of done:
- The comment is posted and the label is moved to agent::needs-review.
- The PM reads it, answers the open questions, and either closes the issue or creates a follow-up build issue.

Agent guardrails: read-only. No merge request, no purchases, no sign-ups.
Labels: agent::ready, type::research

Exercise 3: Build an “Agent Work” board

Open Plan > Issue boards, create a new board called Agent Work, and add one list for each label: agent::ready, agent::in-progress, agent::needs-review (plus agent::blocked if you made it). Add a Closed list at the end. Drag your three issues into ready. Read the docs on issue boards. Agree with your coach on one simple rule: only the PM moves a card into ready, and only the PM moves it out of needs-review.

Exercise 4: Save the review checklist as a template

Copy the checklist below into a merge request description template (ask your coach if you cannot add files; otherwise paste it as the first comment on any agent-made merge request). You review the outcome: does it do what the issue asked, is it safe, and what did it cost? You never need to read code.

## PM review checklist (agent merge request)

Linked issue: #___      Reviewer: ______      Date: ______

### 1. Did it do what we asked? (outcome)
- [ ] Every acceptance criterion from the issue is met (tick each one in the MR)
- [ ] I opened the review app link and tried it myself, on desktop and on a phone
- [ ] The screenshots or recording match what I saw
- [ ] Nothing extra changed that I did not ask for (see "Changes" tab file names for surprises)

### 2. Is it safe to release? (risk)
- [ ] Pipeline is green (tests, security scan, deploy preview)
- [ ] No secrets, passwords or personal data are visible in the description or screenshots
- [ ] The agent stayed inside its guardrails and did not touch out-of-scope areas
- [ ] The MR description explains the change in plain English and lists open questions

### 3. Cost and effort (learning)
- [ ] Number of agent tries: ___   Time to finish: ___   Estimated cost: ___
- [ ] Was the issue clear enough? What would I add next time?

### Decision
- [ ] Approve and hand to an engineer or release owner to merge
- [ ] Request changes (write one clear comment, then set label back to agent::ready)
- [ ] Stop: not what we need (close the MR and note why on the issue)

Exercise 5: Put it on a milestone

Create or reuse a milestone called Agent pilot, assign your three issues, and set a one-week date range. At the end, write a short comment on the milestone: how many issues came back needing changes, and what you would add to the template next time. Docs: milestones.

The agentic angle: you are the agent product owner

This whole page is agentic. An agent is software that repeats a loop: plan what to do, act with its tools (for example, create a file or open a merge request), then observe what happened (test results, errors) before deciding the next step. A product manager does not write that loop, but does own the things that make it work well: the task, the limits, and the review.

The agent loop, with humans in it

  1. You write the task. An agent-ready GitLab issue with acceptance criteria. Label: agent::ready.
  2. HUMAN CHECKPOINT 1 – approve the plan. The agent proposes a plan in a comment. You say go, change it, or stop.
  3. Agent acts. It uses its tools to make changes and opens a merge request. Label: agent::in-progress.
  4. Agent observes. It reads test results and pipeline output, then fixes what failed. This plan, act, observe loop may repeat several times.
  5. Agent stops at its limit. A set number of tries or a budget, then it hands back with notes instead of guessing. Label: agent::needs-review.
  6. HUMAN CHECKPOINT 2 – you review the outcome. Use the review checklist below. Approve, or send it back with a new comment.
  7. HUMAN CHECKPOINT 3 – release. A person merges and launches. Close the issue and note what you learned for the next one.

Tools and MCP

Tools (function calling) are the specific actions an agent may take. MCP is a shared standard for plugging tools and data into an agent. Fewer, clearer tools means fewer surprises.

Sub-agents and orchestration

One agent can hand parts of a job to specialist sub-agents (research, write, test). Someone has to orchestrate and own the result. That someone is you, through the backlog.

Long-running work and checkpoints

Long jobs save progress at checkpoints so they can resume after a failure. Map those checkpoints to your labels and board columns.

Guardrails and permissions

Give an agent the least access it needs. Write guardrails in each issue: what files, systems and spending are off limits.

Evals

An eval is a repeatable test of agent quality. Your acceptance criteria are the first eval: if you cannot test a criterion, rewrite it.

Cost and latency

Each loop costs money and time. Set a limit on tries per issue and compare agent cost with the value of the work.

Teams now run agents inside CI/CD: an agent picks up an issue labelled for it, makes changes on a branch, and opens a merge request for a human to review. GitLab documents its own take in the Duo Agent Platform and flows docs. Ask your coach what is turned on in your workspace before you expect an agent to pick up your issues; the practice above works either way.

Practical exercise: let Claude be your issue reviewer

Paste one of your draft issues into Claude (or another assistant approved by your coach) with this prompt:

You are a strict reviewer of agent-ready issues. Read the issue below. List (1) anything ambiguous, (2) acceptance criteria that cannot be tested, (3) missing guardrails or out-of-scope notes, (4) the three questions a new agent would need to ask before starting. Then rewrite the issue. Do not invent facts about my product; mark gaps with [QUESTION].

Do not paste passwords, customer data or private keys. Answer every [QUESTION] yourself, update the GitLab issue, then ask Claude to attempt the issue as if it were an agent and see where it stalls. Every stall is a missing line in your issue.

Free resource library

Agentic Automation FAQ

Software that is given a goal and repeatedly plans, uses tools to act, and checks the results until it finishes or hits a limit, with people approving the important steps.
No. The no-code track (Hugging Face Unit 1, Zapier, Anthropic sub-agents) needs none. Your GitLab practice is writing, labelling, and reviewing, not coding.
Yes. Every step in the recommended path has a free certificate or a free course. Some tools ask you to make a free account, and n8n exercises need a free trial or a self-hosted copy. The only paid item is the optional DeepLearning.AI Pro membership, marked as later.
Google Cloud Skills Boost offers agent badges (for example, building agents with the Agent Development Kit), but its pages did not confirm they are free, and they are developer-focused. We left them out of the $0 path.
That is the job. Add a comment, correct the acceptance criteria, and move the label back to agent::ready. Repeated misses almost always mean the issue was unclear.

Track your progress

Log each step as a to-do in your GitLab workspace: one issue per step, with the label learning::pilot. Close the issue when you have earned the certificate or finished the course. Five steps, five issues, and your board shows your journey.

Your 12-week PILOT path: Week 10 of 12. Move at your own pace; the order is a suggestion, not a rule.