Phase 2: Core Techniques (Days 4–7) — Build the tools to measure and improve agent output systematically.
Understanding Core Concepts
1. Why the Loop Matters
Without measurement, improvement is guesswork. You change something, run the agent, it "feels" better, so you keep it. But feelings are unreliable—you notice wins and forget losses.
A closed loop captures both: score before, make a change, score after. You see exactly what helped and what didn't.
2. Run → Measure → Improve → Re-measure
This is the cycle:
- Run: Agent processes test cases, produces output
- Measure: Harness scores the output against criteria
- Improve: Based on score, you change instructions, context, or approach
- Re-measure: Run the harness again with the same test cases
Each cycle is data. Over 10 cycles, patterns emerge.
3. Compounding Improvements
Each small improvement sticks. You don't lose progress because you measure it.
Cycle 1: 5/10 cases pass
Cycle 2: 6/10 (improved context)
Cycle 3: 8/10 (tightened spec)
Cycle 4: 9/10 (added constraint)
Progress is visible and compound.
4. The Stopping Point
You run the loop until diminishing returns: the effort to improve is no longer worth the score gain.
For mission-critical features, that might be 95%+ pass rate. For experimental ones, 70% might be enough. The harness tells you where you stand so you can decide.
Tiny Action for Today (20–30 minutes)
Take the harness skeleton from yesterday. Run it.
Record the baseline score. Then:
- Make one small change (tighten a phrase, add a constraint, clarify an example)
- Run the harness again
- Did the score improve, stay the same, or worsen?
Capture the result. This is one closed-loop cycle. You now have proof about what works.