All Posts
AI EngineeringAgent Craft · Part 11 of 12

Day 11 – Using Your Real Project as the Laboratory

Share

Phase 4: Application & Habit (Days 11–12) — Apply what you've learned to your real project and make it stick.

Understanding Core Concepts

1. Why Real Projects Matter

In toy examples, everything is clean. The problem is well-defined, the scope is small, the success criteria are obvious.

Real projects are messy: scope creeps, requirements ambiguous, success is blurry. But that's where you learn. Agent craft isn't a lab exercise—it's for navigating real complexity.

2. Starting Small

You don't rebuild everything at once. Pick one feature the agent is already working on—something you've been iterating on for weeks.

This feature already has:

  • A history of agent mistakes (so you see patterns)
  • Implicit success criteria (you know when it's good)
  • Real context constraints (you know what matters)

Start there.

3. Building Your Harness in Production

Your evaluation harness comes from real failures. Each bug the agent introduced becomes a test case. Each clarification you gave becomes a spec refinement.

The harness isn't written in a vacuum—it's built from experience.

4. The Feedback Loop Accelerates

With a real project, feedback is immediate. The agent makes a change, it runs in your app, users (or your tests) catch issues.

That feedback tightens the loop. You iterate faster because you see impact directly.

Tiny Action for Today (30–45 minutes)

Pick one feature in your real codebase the agent has touched. Do this:

  1. Collect recent attempts: Find 5 versions the agent produced (errors, incomplete work, revisions)
  2. Identify patterns: What went wrong across these attempts?
  3. Write it down: Create a one-paragraph spec for how this feature should work based on what you've learned
  4. Create a test: Write 2–3 concrete test cases from real failures

This is your harness skeleton, built from production experience, not theory.

Share
Back to all posts