AI Agent Design Patterns

Each pattern directory contains an interview.md with 7–8 questions spanning four categories: Conceptual → Trade-offs → Implementation & Failure Modes → Extension

Questions are ordered from definition-level to system-design depth, mirroring how interviewers typically probe a topic.

---

Study Path (simple → complex)

Foundational Patterns

These establish the vocabulary. Learn these first.

#	Pattern	Core idea	Interview file
02	Routing	Single-classifier dispatch to specialized handlers	patterns/02-routing/interview.md
01	Prompt Chaining	Fixed sequential LLM pipeline	patterns/01-prompt-chaining/interview.md
03	Parallelization	Fan-out to N workers, fan-in to aggregator	patterns/03-parallelization/interview.md
06	Code Execution	Generate → execute → retry with error feedback	patterns/06-code-execution/interview.md

Iterative Loop Patterns

These introduce cycles and self-improvement.

#	Pattern	Core idea	Interview file
07	ReAct	Reason-Act-Observe loop with tool calls	patterns/07-react/interview.md
08	Reflection	Generate → self-critique → revise until stable	patterns/08-reflection/interview.md
05	Evaluator-Optimizer	Separate generator and evaluator in a revision loop	patterns/05-evaluator-optimizer/interview.md
16	Constitutional AI	Critique-revision loop guided by explicit principles	patterns/16-constitutional/interview.md

Planning & Decomposition Patterns

These split hard problems into manageable parts.

#	Pattern	Core idea	Interview file
12	Self-Ask	Decompose into sub-questions, answer sequentially	patterns/12-self-ask/interview.md
20	Least-to-Most	Order sub-problems easiest-first, build up context	patterns/20-least-to-most/interview.md
09	Plan-and-Execute	Static plan → per-step inner ReAct → synthesize	patterns/09-plan-and-execute/interview.md
18	Speculative Execution	Generate N independent candidates, evaluate, pick best	patterns/18-speculative/interview.md

Multi-Agent & Coordination Patterns

These involve multiple models or structured control flow.

#	Pattern	Core idea	Interview file
04	Orchestrator-Workers	Plan + isolated workers + synthesis	patterns/04-orchestrator-workers/interview.md
10	Multi-Agent	Supervisor routes to role-based agents, synthesizes	patterns/10-multi-agent/interview.md
17	Mixture-of-Experts	Router selects top-k expert agents to answer	patterns/17-mixture-of-experts/interview.md
15	Debate	Affirmative vs. Negative agents, judge synthesizes	patterns/15-debate/interview.md

Stateful & Constrained Patterns

These add persistent state, explicit structure, or human oversight.

#	Pattern	Core idea	Interview file
11	Memory-Augmented Agent	ReAct + persistent MemoryStore (remember/recall/forget)	patterns/11-memory/interview.md
14	State Machine	Explicit state graph; LLM selects transitions	patterns/14-state-machine/interview.md
19	Event-Driven	Per-event mini-ReAct loop with persistent state	patterns/19-event-driven/interview.md
13	Human-in-the-Loop	ReAct with checkpointed tools requiring human approval	patterns/13-human-in-the-loop/interview.md

---

Cross-Pattern Comparison Questions

These are high-signal interview questions that span multiple patterns:

1. ReAct vs. Plan-and-Execute: When does dynamic replanning (ReAct) outperform a static upfront plan, and vice versa?
2. Reflection vs. Evaluator-Optimizer: Both use a critique step — what does using a separate evaluator model add, and at what cost?
3. Routing vs. Mixture-of-Experts: Both dispatch to specialists — how do they differ in how many specialists answer, and when does each model fit better?
4. Orchestrator-Workers vs. Multi-Agent: Both decompose work across agents — what determines which pattern to use for a given task?
5. Self-Ask vs. Least-to-Most: Both decompose into sub-questions — what does ordering by difficulty add, and when does it not matter?
6. Parallelization vs. Debate: Both generate multiple answers — how does the adversarial rebuttal structure in Debate add value that parallel independent answers do not?
7. State Machine vs. ReAct: Both loop until a terminal condition — what does making the control flow graph explicit gain you, and what does it cost?
8. Speculative Execution vs. Evaluator-Optimizer: Both try multiple approaches and select the best — what is the fundamental difference in how they improve quality?