Context & Reliability — Glossary

Context Management & Reliability · 15% of the exam. 10 terms.

The finite token budget available for a request — input history and output share this same budget.

Exam context: Budgeting questions expect you to account for tool definitions and retrieved data, not just the conversation text.

Marking a stable prefix of a prompt as cacheable so repeated requests reuse it server-side, cutting latency and cost.

Exam context: Only the portion before the first change is reusable — appending new content after a cached prefix keeps the cache hit.

Deliberately allocating context space between system prompt, tool definitions, retrieved data, and history so nothing critical is crowded out.

Exam context: The exam rewards proactively trimming low-value context over reactively hitting the limit and truncating blindly.

Condensing older conversation history into a summary to free context space.

Exam context: The 'summarization trap': over-compacting drops details the model still needs, producing vague, ungrounded later responses.

Surfacing uncertainty to a human rather than silently guessing when a decision is high-stakes or underspecified.

Exam context: Correct answers favor escalation exactly when reversibility is low and confidence is uncertain — not for every judgment call.

Deciding whether a subagent's error should halt the parent process, trigger a retry, or surface as a degraded-but-continuing result.

Exam context: Silently swallowing every subagent error is as wrong as halting on every one — the right call depends on the task's criticality.

Designing a tool call so running it twice has the same effect as running it once.

Exam context: Essential for anything that might get automatically retried — non-idempotent actions (like sending an email) need a guard.

An agent's strategy for building situational context — grep/search first vs. reading whole files — before making changes.

Exam context: Exam scenarios reward targeted exploration over reflexively reading every file in a large repo.

A structured test suite scoring agent or prompt outputs against expected behavior, used to catch regressions before shipping.

Exam context: Treat evals as the mechanism that turns 'it seemed to work when I tried it' into something you can actually trust.

Tracking where a piece of context came from — which tool call, file, or turn — so it can be trusted appropriately.

Exam context: Matters most when a model must reconcile conflicting information from two different sources in the same context.