Domain 5 · Context Management & Reliability

Reliability & escalation

A reliable agent isn't one that always succeeds — it's one that knows its limits and hands off cleanly when it hits them. The skill the exam tests is designing escalation triggers and stopping error propagation, so a single failure degrades gracefully instead of cascading into a confident wrong answer.

escalation triggerserror propagation graceful degradationhuman handoff confidence

Explain like I'm 10

A good lifeguard isn't the one who never needs help — it's the one who blows the whistle and calls for backup the moment something's beyond them, instead of pretending everything's fine. An agent should do the same: when it's unsure, when the stakes are high, or when a tool keeps failing, it should escalate to a human — not bluff its way through and make things worse.

Decide: handle, retry, or escalate?

Pick a situation and choose the agent's move. The reliable answer often isn't "try harder" — it's "stop and escalate." See which trigger fires:

situation

Escalation triggers worth wiring in

Trigger	Why escalate
Exceeds authority	Action is beyond the agent's policy limits (refund cap, account deletion).
Low confidence / ambiguity	The agent isn't sure — better to ask than to guess on something that matters.
Repeated tool failure	A dependency is down; retrying forever wastes turns and frustrates the user.
High emotional / legal stakes	Threats, safety, legal, or vulnerable-user signals need a human.
Loop / budget limit hit	The guardrail tripped (Domain 1) — hand off rather than fail silently.

Stopping error propagation

In a multi-step or multi-agent flow, a wrong/missing result from step 2 must not be silently consumed by step 5 as if it were valid. Propagate status alongside data: mark a result partial or failed, and have downstream steps check it. The synthesis must account for the gap — degrade explicitly ("I couldn't verify X") rather than presenting an incomplete result as complete.

Exam trap: "keep retrying until it works" and "give the best guess anyway" are both wrong when the situation calls for a human. Reliability = explicit escalation on the right triggers and visible degradation, never a confident answer built on a failed/partial step. Silent failure is the cardinal sin.

Takeaways: reliable agents escalate on clear triggers — exceeds authority, low confidence, repeated failure, high stakes, or a tripped guardrail — instead of bluffing. Stop error propagation by carrying status (partial/failed) with data and degrading explicitly. A clean handoff beats a confident wrong answer every time.

Curated companion: Anthropic — Building Effective Agents (human oversight).