A city CIO wants civic Claude agents that are operable and auditable rather than ad-hoc chat scripts. Which construction preference supports that goal?

Preferring Claude Agent SDK patterns over ad-hoc chat loops supports a city CIO goal of civic agents that are operable and auditable rather than ad-hoc chat scripts. SDK-structured construction yields clearer control flow, versioned tool registration, hooks, and operational touchpoints that change management and auditors can review. Ad-hoc chat loops are napkin sketches; Agent SDK patterns are stamped blueprints cities can operate across environments and staff turnover.

Designing civic workflows around interactive Claude chat threads for rapid iteration and review fails by concept: it treats agents as conversational scratchpads rather than versioned, reviewable operational systems, leaving auditability dependent on ephemeral chat history. Configuring agent tool access through Model Context Protocol servers for live civic data feeds addresses data connectivity, not the construction preference for auditable, structured agent control flow; it does not replace the need for SDK-defined agents over chat loops. Optimizing civic agent latency with prompt caching across repeated context blocks in production is a performance tuning concern, not an agent construction preference; it does not make ad-hoc chat scripts operable or auditable.

Exam caveat: using the SDK is necessary but not sufficient for auditability—you still need logging, access control, tests, and promotion pipelines. Do not equate "has SDK" with "is governed."

Operational check: implement agents via Agent SDK with version-controlled configs, peer-reviewed PRs, secrets in a vault, do not rely on prompt caching for auditability, enable structured logs/hooks, and require staging sign-off before municipal production rollout.