Learning: Single Agent Workflows¶
How to Use This Page¶
- Keep each loop step observable.
- Stop on evidence, not intuition.
- Track both quality and cost per run.
Lesson 1: Fundamentals¶
Learn the plan -> act -> observe loop and why explicit loop state matters.
What You Learn¶
- How to separate planning from action execution
- Why observation logs improve final answer quality
- How step limits prevent runaway loops
Walkthrough¶
- Build a loop with explicit plan, action, and observation fields.
- Run for a fixed maximum step count.
- Produce a final answer from accumulated evidence.
Try It Yourself¶
- Remove observation tracking and compare output quality.
- Test step limits of 3, 5, and 8.
- Add a guard for repeated actions.
Lesson 2: Intermediate Patterns¶
Learn evidence sufficiency checks and fallback behavior.
What You Learn¶
- How to determine when evidence is "enough"
- Why repeated low-yield steps should trigger fallback
- How confidence notes improve answer transparency
Walkthrough¶
- Add an evidence scoring rule.
- Stop when score crosses threshold.
- Route low-score runs to fallback response logic.
Try It Yourself¶
- Compare strict vs lenient evidence thresholds.
- Add a fallback that requests clarification from the user.
- Log how often fallback is triggered.
Lesson 3: Production Patterns¶
Learn adaptive planning and information-gain-driven stopping.
What You Learn¶
- How to estimate per-step information gain
- Why early stopping can reduce cost without quality loss
- How to escalate uncertain results safely
Walkthrough¶
- Score expected gain for candidate next actions.
- Select highest-gain action first.
- Stop when expected gain falls below threshold.
Try It Yourself¶
- Compare fixed-plan vs adaptive-plan loop performance.
- Measure cost savings from early stopping.
- Add an escalation path for uncertain final outputs.
Code Examples¶
- Starter script:
docs/starter/stage05_single_agent.py - Stage architecture notes:
docs/roadmap/stage-05-planner-executor-loops/architecture.md