6 · Agentic AI

Subtitles

1. Tool requests and the agent harness
2. ReAct: reason, act, observe
3. Predefined workflows versus agents
4. Prompt chaining, routing, and parallelization
5. Orchestrator-workers and evaluator-optimizer
6. Stopping, checks, and human approval
7. Trajectory, outcome, and cost evaluation
8. AI-SCE and clinical transfer

Milestones

• Trace a tool request through execution and observation.
• Classify every workflow pattern from its control flow.
• Justify an agent only when paths depend on observations.
• Define acceptance, revision, stopping, and escalation criteria.
• Evaluate trajectories, outcomes, cost, and clinical transfer.

Possible exam questions

OPEN · Q1 — A model writes correctly formatted tool requests. Is tool integration therefore complete? What else must be established?

OPEN · Q2 — All analysis paths and exceptions are known in advance. Choose a predefined workflow or an agent, based on what determines the next step.

OPEN · Q3 — Should independent record summaries use parallelization, and a case where findings determine the next review use orchestrator-workers? Why?

OPEN · Q4 — Design evaluator–optimizer acceptance and stopping criteria for summaries that must preserve findings, values, units, and chronology.

OPEN · Q5 — Design an AI-SCE comparison of a fixed record-review workflow and an agent, covering outcomes, process, interactions, and resource use.

RECALL · Q6 — In a ReAct-style system, where does the observation after a search request come from?

RECALL · Q7 — What defines the ReAct approach?

RECALL · Q8 — How do few-shot demonstrations influence a ReAct model without changing its weights?

RECALL · Q9 — Why can external search reduce factual hallucinations?

RECALL · Q10 — What is the main distinction between workflows and agents?

RECALL · Q11 — Extract findings → draft summary → translate summary: which workflow pattern is this?

RECALL · Q12 — Classify a request and send it to a specialist handler: which workflow pattern is this?

RECALL · Q13 — Which agent capability is least directly tested by a static medical multiple-choice benchmark?

RECALL · Q14 — What is the role of agent-based modeling in clinical evaluations?

RECALL · Q15 — After strong simulated performance, what evidence is needed for transfer to actual care?