Skip to content

Phase 3 — Agentic Systems

Goal: Build sophisticated agent workflows similar to those expected in modern enterprise AI deployments.

Master

  • LangGraph (secondary/comparative)
  • ADK (priority)
  • ReAct
  • Tool Use
  • Agent Planning
  • Multi-Agent Architectures
  • State Management

Self-Interview Check

Ask each question cold and out loud (~2–3 min each). Score honestly: 1 = couldn't answer · 3 = partial / hand-wavy · 5 = crisp with production depth. Cross-references like §5.1 point to deep dives in the other phase docs.

Core Patterns

  1. "Explain the ReAct loop. When does reasoning + acting beat a single well-prompted LLM call — and when is it overkill?"
  2. Strong answer: multi-step tasks needing observations between steps; honestly states a fixed chain or single call often wins on cost, latency, and reliability.
  3. "Design a tool-using agent that looks up operational data via an API. How do you define the tools, validate calls, and handle tool failures?"
  4. Strong answer: tight typed tool schemas, argument validation, retry vs. escalate decisions, clear error messages fed back to the model.
  5. "What stops an agent from looping forever, spiraling on tool errors, or burning budget?"
  6. Strong answer: max iterations, tool-call/token budgets, circuit breakers, human checkpoints before irreversible actions.

Frameworks & Architecture

  1. "How do ADK and LangGraph each model agents and control flow? Which would you pick for a GCP deployment and why?"
  2. Strong answer: honest comparison — ADK's agent/tool model + Vertex integration vs. LangGraph's explicit state graphs; picks based on fit, not hype.
  3. "When is a multi-agent system justified over one agent with more tools?"
  4. Strong answer: specialization, context isolation, parallel work — weighed against coordination cost, debuggability, and latency. Default stays single-agent.
  5. "Walk through a hierarchical/orchestrator-worker agent design for a real task. Who holds the plan? How do results merge?"
  6. Strong answer: clear division of planning vs. execution, how sub-agent outputs are validated/merged, failure propagation.

State, Memory & Testing

  1. "How do you manage state and memory across a long-running, multi-step agent task — including resume after a crash?"
  2. Strong answer: checkpointing/persisted state, session memory, summarization for long contexts, idempotent resumption.
  3. "How do you test and evaluate a non-deterministic agentic workflow?"
  4. Strong answer: trajectory evals, assertions on tool calls and outcomes, trace replay, scenario eval sets — not just "ran it and looked right".
  5. "An ops director says: 'build me an AI agent to fix our task handoffs.' What do you verify before writing any agent code?" (see Phase 5, §5.1)
  6. Strong answer: discovery first — the fix might be a filtered list, not an agent; only builds agentics after validating a multi-step reasoning need.

Phase Gate

  • Scored 4+ on ≥80% of the questions above, out loud, without notes.
  • First ADK agent running.

Notes / Evidence