Domain 1 · Lesson 3 · Agentic Architecture (27%)

Enforcement, Hooks & Handoffs

Task Statements 1.4 & 1.5 — the most heavily tested idea on the whole exam.

Course progress: Domain 1 ▸ lesson 3 of ~4

If you learn one thing for the exam, learn this: when you need a guarantee, use programmatic enforcement — not a prompt. Prompts have a non-zero failure rate. Hooks and prerequisite gates do not.

Programmatic enforcement vs prompt-based guidance

Prompt-basedProgrammatic (hooks / gates)
NatureProbabilistic — the model usually compliesDeterministic — the rule always fires
Use whenGuidance, defaults, soft preferencesA business/safety rule must be guaranteed
Example"Always verify the customer first"Block process_refund until get_customer returned a verified ID
The load-bearing rule Identity verification before financial operations, refund caps, compliance ordering — anything where an error has real consequences → enforce in code. "Enhance the system prompt to say it's mandatory" and "add few-shot examples showing the right order" are distractors when the stem asks for a reliability guarantee.

Prerequisite gates

A prerequisite gate blocks a downstream tool until an upstream step has completed. Canonical exam case: the agent sometimes skips get_customer and calls lookup_order on a stated name, misidentifying accounts. The fix that works: a programmatic prerequisite that blocks lookup_order and process_refund until get_customer has returned a verified customer ID. Prompts and few-shot examples reduce the rate but can't drive it to zero; a routing classifier addresses tool availability, not tool ordering — wrong problem.

Hooks: intercepting the loop

Hooks are code that runs at defined points in the agentic loop. Two patterns the exam names:

Why hooks over prompts hereBoth patterns are chosen precisely because they are deterministic. "Tell the model in the prompt not to refund over $500" is the probabilistic distractor.

Multi-concern decomposition

When one customer message contains several issues (a refund and a billing dispute and an address change), decompose it into distinct items, investigate each — in parallel where possible using shared context — then synthesize one unified resolution. Don't tunnel on the first issue.

Structured handoff on escalation

When escalating to a human who has no access to the conversation transcript, hand off a structured summary, not "customer is upset." Include: customer ID, root-cause analysis, refund amount / relevant figures, and a recommended action. The human should be able to act without re-interviewing the customer.

Check yourself

In 12% of cases the agent skips customer verification and refunds the wrong account. Which change most effectively guarantees the reliability fix?
Correct: option 3. A required sequence protecting money demands a deterministic guarantee. Prompt enhancement and few-shot (options 1–2) are probabilistic and retain a non-zero failure rate. The classifier (option 4) changes which tools are available, not the order they run in — it solves a different problem.
Different MCP tools return dates as Unix time, ISO 8601, and numeric codes, confusing the agent. The cleanest place to normalize before the model reasons over them is a:
Correct: option 1. PostToolUse intercepts results before the model processes them — exactly where deterministic normalization belongs. A prompt (option 2) is probabilistic; a post-hoc subagent (3) runs too late; forcing tool order (4) doesn't reconcile formats.
Decision rules to carry forward Guarantee needed → hook or prerequisite gate (deterministic). Guidance/default → prompt. Normalize tool data → PostToolUse. Block policy-violating call → interception hook. Human handoff → structured summary (ID, root cause, amount, recommended action).
Ask your teacher. Want the full list of Agent SDK hook types, or how a gate is actually wired vs an interception hook? Ask before Lesson 4.