Phase 3: Systems & Production (Days 8–10) — Scale evaluation to real projects and production constraints.
Understanding Core Concepts
1. The Autonomy Paradox
Tight control (detailed step-by-step instructions) produces consistent results but they're often suboptimal—the agent can't adapt to surprises or find better paths.
Loose control (just a goal, no guardrails) produces creative results but unreliable ones—the agent wanders.
The sweet spot is: clear goal, loose path, hard constraints.
2. Bounds vs Paths
A path tells the agent how to do something step-by-step:
- "First parse the JSON, then iterate through the keys, then..."
Bounds tell the agent what it can't do:
- "Don't modify the database schema. Keep the API contract. Security first."
Bounds give the agent room to explore while preventing damage.
3. Constraints as Freedom
This sounds backwards, but constraints enable speed. The agent spends less time wondering "is this allowed?" and more time solving the problem.
Good constraints are:
- Clear (unambiguous)
- Justified (the agent understands why)
- Minimal (only what truly matters)
4. Observability Inside the Bounds
When the agent has autonomy, you need to see what it's doing. This is why evaluation and logging matter.
You set bounds, let it work, then measure. If it consistently hits the same constraint, that constraint wasn't the right one. Adjust and try again.
Tiny Action for Today (15–20 minutes)
Think of a task you've given an agent recently. Write down:
- What was the goal?
- What constraints did you set?
- Did those constraints prevent problems or create unnecessary friction?
Then rewrite the instructions as pure bounds (no step-by-step paths). Remove prescriptive language, add only protective guardrails.