Capability bloat
An oversized tool or agent surface — unused, overlapping, or role-mixed capabilities — that raises mis-selection risk and blast radius.
Exam context: Fix by remove/merge/split before longer prompts or more tools.
See also: 3.1 Capability bloat
Least privilege
Grant each agent, credential, and tool only the actions and resources required for its role; remove unused capability rather than merely logging it.
Exam context: Least privilege removes unused tools — confirmation dialogs are not a substitute.
See also: 3.1 Capability bloat
Tool scope
The bounded set of tools (and permissions) attached to an agent role. Narrow scopes improve selection and security reviews.
Exam context: Split write/admin tools off customer-facing read agents when scopes collide.
See also: 3.1 Capability bloat
AuthN vs AuthZ
Authentication verifies identity; authorization decides what that identity may do to which resource. Both are required on agent tool paths.
Exam context: If login works but admin tools still run, the gap is AuthZ.
See also: 3.2 Auth security gaps
Accuracy–latency trade-off
Configuration choices (model, retrieval depth, tool fan-out) move quality and speed together; architects quantify both against an SLA.
Exam context: Accuracy gains that break a hard p95 without a waiver are no-ship.
See also: 3.3 Accuracy–latency
Observability sampling
At high volume, keep aggregates always-on and sample detailed traces with boosts on errors, outliers, and canary failures — with redaction.
Exam context: Prefer strategic sampling over tracing everything or tracing nothing.
See also: 3.4 Observability at scale
Chunking
Splitting source documents into retrieval units. Structure-aware boundaries (clauses, Q+A, tables) beat blind fixed token windows.
Exam context: One-size chunks that break clause meaning are a classic trap.
See also: 3.5 RAG chunking & indexing
Embedding index
The searchable vector (and often hybrid) store built from chunk embeddings. Upload to object storage is not the same as a live index.
Exam context: After a corpus refresh, verify re-embed and index promote before blaming the model.
See also: 3.5 RAG chunking & indexing
Provenance metadata
Fields on chunks and citations (source URI, section, version, date, hash) that enable filters, audits, and grounding checks.
Exam context: Metadata supports filtering and provenance — not optional decoration.
See also: 3.5 RAG chunking & indexing
Hybrid retrieval
Combining sparse/keyword and dense semantic signals (often with structured filters) so exact tokens and paraphrases both rank well.
Exam context: SKU/error-code traffic that fails under dense-only usually needs hybrid or sparse.
See also: 3.6 Retrieval strategies
MCP (Model Context Protocol)
A shared protocol for exposing tools and resources to hosts/clients so multiple products can reuse one integration surface.
Exam context: High reuse of tools/resources → MCP; narrow owned path → API/CLI may suffice.
See also: 3.7 Connection protocols
Agent-to-agent
Integration pattern where specialist agents communicate via explicit handoffs, keeping separate contexts and tool scopes.
Exam context: Use for role isolation — not as a substitute for a shared tool catalog.
See also: 3.7 Connection protocols
Progressive discovery
Load only the files, chunks, or symbols needed for the next step (search → open → extract → summarize) instead of dumping a whole corpus.
Exam context: Default for multi-GB or unknown corpora that cannot fit.
See also: 3.8 Progressive vs monolithic
Monolithic context
Packing a bounded, usually small and stable corpus into the prompt (often cached). Requires an honest token budget and headroom.
Exam context: Fine for tiny stable packs; wrong as an every-turn strategy for huge repos.
See also: 3.8 Progressive vs monolithic
Stable prefix caching
Caching unchanging system/tool instruction prefixes to cut latency and cost on repeated calls without changing behavior.
Exam context: Often the first latency win before upgrading model tier when accuracy already clears the floor.
See also: 3.3 Accuracy–latency
Confused deputy (agent tools)
Risk where an authenticated agent uses overly broad credentials or free-form IDs to act beyond the end-user’s entitlements.
Exam context: Bind actions to verified session context and scoped credentials.
See also: 3.2 Auth security gaps