A county IT lead must choose Anthropic-hosted managed agents versus self-hosting the agent runtime inside the county VPC. Which factor most strongly favors self-hosting?

A requirement that the agent runtime stay inside municipal network boundaries most strongly favors self-hosting the agent runtime inside the county VPC versus Anthropic-hosted managed agents. Data residency and network-boundary policies often mandate that case-handling runtimes, connectors, and logs remain on county-controlled infrastructure. Self-hosting aligns the Claude agent runtime with those controls when managed hosting would place execution outside the approved boundary, even if model inference options differ by architecture.

Desire for vendor-managed scaling, patching, and monitoring of agent workloads fails as a self-hosting driver because those are operational tasks the county takes on itself when running the runtime in its own VPC. Need for centralized billing and usage metering via hosted APIs favors managed hosting rather than self-hosting, so it does not support placing the runtime inside the county VPC. Preference for automatic model upgrades handled by the hosted platform argues for managed agents, not self-hosting, because the county would lose that hands-off update model when it runs the runtime itself.

Exam caveat: self-hosting favors boundary control but increases operational burden (patching, scaling, monitoring). Managed agents may still be acceptable when policy allows and data-handling agreements meet county requirements—read the residency clause carefully.

Operational check: map which runtime components must remain in-VPC, confirm network egress rules for model calls if any, document the residency decision against county policy, and validate that case payloads and agent logs do not land in disallowed external runtimes.