Four weeks from user to engineer

The plan · weeks 1 to 4

One harness skill per week

• Week 1: context. Cut your AGENTS.md or CLAUDE.md to what every task needs; move one long procedure into a skill. Compare a task before and after.
• Week 1: first sensor. Add one blocking hook for your most expensive repeated mistake and watch it fire.
• Week 2: observability. Record time, tokens, cost and outcome per agent task. Compute cost per accepted task for one week of work.
• Week 3: a goal loop. Run one task as a loop with a verifiable exit, an iteration cap, a spend cap and a no-progress check, in its own branch.
• Week 4: evals. Build a 20-case golden set from real tasks, record a baseline, then rerun it after one harness change. Report pass rate over several trials.
• Week 4: share it. Write up one failure and the harness change that fixed it, with before and after numbers.
• All month: flashcards on schedule until every card reaches the last box.

Habits

What experienced agentic engineers do differently

• Read the transcript

  When a run goes wrong, read what the agent actually saw before blaming the model.

• Fix the harness, not the run

  Every repeated failure becomes a guide, a sensor or a permission.

• Measure before changing

  No prompt or harness change without a before and after.

• Verify, don't trust

  "Done" means a check passed, not that the agent said so.

Sources

Where the claims in this course come from

Terms in this field are young. Every definition here traces to one of these sources; read them for depth, and expect the vocabulary to keep moving.