The Agent Runtime SDK is enterprise template work: a production-grade foundation for building AI agents that survive their own infrastructure decisions. Its core promise is re-routability on three layers. Model providers, execution runtimes, and distribution channels can each be swapped without touching agent code, and a migration drill in the repo proves the promise.
Each agent is a portable, schema-validated manifest: identity, instructions, model routing, tools, skills, and fine-grained permissions in one document. Governance runs on two planes, who may invoke the agent and what the agent may see on the user’s behalf, both enforced at the gateway to shut down confused-deputy attacks. Multi-tenant provisioning mints per-agent API keys with hard cost budgets and timeouts; tokens are never persisted.
It ships as infrastructure: multi-cloud Terraform, incident runbooks, secret scanning in CI, and OpenTelemetry through every layer. It grew out of the base-plate stack that carries Weapp’s enterprise engagements.