Teaching agentic engineering

Use this guide when you are mentoring one person, running a workshop, reviewing a candidate, or helping a team adopt agentic engineering.

The goal is simple: help the learner turn a real problem into agent-directed work they can scope, review, prove, and hand off.

Start here

If the room is mixed, start smaller than you think. A good first exercise is one file, one command, one fixture, or one failing test.

The loop you are teaching

Teach the same loop every time. The tool can change. The behavior should not.

Ready-to-run agendas

30-minute mentoring

Use this when one learner is stuck or trying the workflow for the first time.

90-minute team session

Use this when a team needs a shared baseline.

1-day in-person workshop

Use this when people are coming from different backgrounds. Keep the morning shared. Make the afternoon role-specific.

Shared lab default

Use Data + AI practice labs when the learner does not bring a project.

Start every workshop with the same local shape:

mkdir agentic-lab
cd agentic-lab
git init
mkdir -p data lab-artifacts

Then ask the learner to create one file before using the agent:

cat > lab-artifacts/brief.md <<'EOF'
# Brief

Goal:
Constraints:
Inputs:
Output:
Done check:
Risk to review:
EOF

What learners submit

Every learner should finish with a small artifact package:

• brief.md: goal, constraints, inputs, outputs, done check, and risk.
• plan.md: the plan they approved or changed.
• proof.md: command output, screenshot notes, eval result, logs, or deployed URL.
• decision.md: what changed, what was rejected, and why.
• next.md: the next smallest slice another person could run.

That package matters more than a perfect final product. It shows whether the learner can direct the work.

The five instructor moves

1. Start with the smallest real thing

Do not let the learner begin with a platform rewrite. Pick one command, one route, one report, one broken test, one data fixture, or one support workflow. The work needs to be real enough to matter and small enough to finish.

2. Make the learner say the done check first

Before the agent edits files, ask:

• What output should exist?
• What command proves it?
• What failure would make you reject the result?
• What risk needs human review?

If the learner cannot answer, help them shrink the slice.

3. Stop at the diff

The most important classroom moment is not the prompt. It is the diff. Ask the learner to explain:

• what changed,
• why it changed,
• what they would keep,
• what they would reject,
• what proof exists.

If they cannot explain the diff, they do not own the work yet.

4. Teach tool choice as translation

Use Claude Code as the main path because the course examples are written that way. Then translate the habit:

• Claude Code: plan mode, CLAUDE.md, permissions, subagents.
• Codex CLI: AGENTS.md, approval modes, local command review, code review agent.
• Gemini CLI: large-context exploration, file tools, shell tools, web fetch, memory, trusted folders.
• Coco: governed data workspace, approved data actions, enterprise evidence.

The habit is stable. The command surface changes.

5. Finish with a handoff artifact

Every workshop or cohort should end with a small package:

• brief,
• plan,
• diff,
• proof command output,
• eval or test,
• decision note,
• next-slice instruction.

That is the smallest version of FDE behavior.

Instructor checklist

Before the session:

• Pick the shared lab or learner project.
• Prepare one local fixture so nobody is blocked by credentials.
• Decide the primary CLI and one fallback CLI.
• Write the proof command on the board.
• Open the relevant course page and lab page.

During the session:

• Keep slices small.
• Make learners read diffs.
• Reject vague “it works” claims.
• Ask for command output, screenshots, logs, or test results.
• Capture the next instruction before moving on.

After the session:

• Ask each learner to commit the artifact package.
• Have them write one paragraph on what the agent got wrong.
• Have them write the next slice as a brief.
• For teams, turn the best artifact into AGENTS.md, HANDOFF.md, or a team playbook.

Reference layer

Use these as instructor anchors, not as extra reading homework.

Common failure modes

What good sounds like

Use this language in the room:

• “Show me the command that proves it.”
• “What did the agent change?”
• “What would make you reject this?”
• “What is the next smallest slice?”
• “Where will future-you read this decision?”
• “Which part is workflow, which part is agent?”

That language is the curriculum.