Domain 3: Integration

Cheat sheet for Integration: bloat, auth, latency, observability, RAG, retrieval, protocols, context strategy.

Tool / agent capability bloat (3.1)

Shrink the choice set before rewriting prompts. Inventory → merge duplicates → remove unused → split write/admin by role → re-eval selection.

Overlapping names = selection noise
Zero-call pilot tools → delete
Least privilege removes capability; logging/confirm does not
One mega-agent with every integration is usually wrong

Auth gaps (3.2)

AuthN = who · AuthZ = what on which resource
Enforce AuthZ in tool adapter / API, not only prompts
No god-mode shared service accounts
Bind mutations to verified session/case context
Fail closed — do not escalate privilege silently

Accuracy vs latency (3.3)

State accuracy floor + p95/cost SLA. Quantify Δ for every change. If floor is met, prefer faster/cheaper. Cache stable prefixes before buying a bigger model.

SLA-breaking accuracy bump without waiver → redesign or reject
Profile retrieval hops and cold prefixes first
Interactive vs batch paths can use different configs

Observability at scale (3.4)

SLIs: quality/canary, latency, cost, tool errors — not only uptime
Aggregates always-on; sample detailed traces; boost on errors
Redact secrets; join request_id to outcomes
Alerts → runbooks → eval fixtures

RAG chunking & indexing (3.5)

Structure-aware chunks (clauses, Q+A, tables)
Metadata: source, section, version, date, hash
Refresh = re-chunk + re-embed + smoke + promote
Wrong after refresh → check retrieval/index before model swap

Retrieval matching (3.6)

Exact IDs → sparse/hybrid (+ filters). Paraphrase → dense. Faceted browse → filter then rank. Score per query bucket, not only macro averages.

Protocol choice (3.7)

High reuse tools/resources → MCP (shared server)
Owned narrow single consumer → API/CLI
Role isolation → agent-to-agent + explicit handoffs
Never: secrets in prompts as integration

Progressive vs monolithic (3.8)

Large / unknown / changing → progressive discovery
Small stable pack that fits → monolithic (+ cache)
Pin facts (IDs, constraints) outside rolling summaries
Multi-GB every-turn dump is never the answer