Agent Ops: Protocols, Gateways, and Production Harnesses

How building software is changing at Anthropic. Anthropic retools engineering so agents and AI handle code review, testing, and production agent infrastructure, reshaping team structure and workflows. This is a field-level example of agentic teams and orchestration in practice — takeaway for Principle 03 and 09: reorganize roles, CI, and ownership around agent loops.

Agentic AI Foundation updates MCP with stateless architecture, hardened authentication, and 12-month deprecation policy. The Model Context Protocol revamp tightens authentication, makes context stateless, and formalizes deprecation timelines for agent integrations. Outcome engineers must update context-handling and lifecycle plans to avoid brittle integrations and keep agent chains auditable — connects to Principle 06 and 10.

Dynamic Workflows: the feature that enabled the Bun rewrite. Pydantic’s DynamicWorkflow lets models orchestrate swarms of agents and provides auditable, large-scale multi-agent code workflows. This gives a practical library-level primitive for building observable, testable orchestration — directly relevant to Principle 09 and 13 for building audit trails.

Snowflake launches Cortex AI Gateway to control AI agents and prevent runaway enterprise costs. Cortex AI Gateway centralizes agent governance, data access control, and cost throttling across agent fleets. If you run agents in enterprise contexts, a gateway or control plane like this is now essential for identity, policy enforcement, and cost containment — see Principle 10 and 15.

Visa used Mythos to hunt for bugs in its own payment network, then open-sourced the harness that made it possible. Visa ran Anthropic’s Mythos to find deep exploit chains and published the agentic harness they used to govern security hunts. Use that harness as a reference pattern for building governed, auditable agent tooling and red-team pipelines — relevant to Principle 07 and 14.