Designing Portable Agents presents a practical architecture for building provider-neutral agent systems in TypeScript using open standards and explicit trust boundaries. Rather than treating an agent as a prompt wrapped around a model, the book develops it as a bounded software system composed of model adapters, typed tools, reviewed skills, document-ingestion pipelines, evidence controls, durable state, and observable execution.

A running research workflow demonstrates the architecture end to end: the agent admits documents, spreadsheets, and images; discovers and reads public sources; performs isolated computation; coordinates multiple model providers; and produces a verified presentation artifact. Each boundary is examined in detail, including strict declarative configuration, streaming and termination, tool authorization, retrieval and citation provenance, capability negotiation, fallback, retries, cancellation, effect reconciliation, and replay.

The book separates interface portability from semantic equivalence. Shared APIs can normalize requests and events, but they do not guarantee identical provider behavior. Portability therefore depends on explicit capability evidence, disclosed recipients, task-specific fallback contracts, and tests tied to exact implementation revisions. Security is treated as a system property: retrieved and generated content remains untrusted, tools grant only server-owned authority, and production deployment requires identity, isolation, egress control, redaction, evaluation, and rollback gates.

The final chapters connect these patterns to OpenTelemetry, MCP, A2A, and the Open Responses specification while carefully distinguishing architectural resemblance, compatible-subset behavior, official test results, and formal conformance. The result is a reproducible guide to constructing agents that can change models and providers without surrendering control over authority, evidence, execution, or operational risk.