Agent infrastructure: MCP, gateways, harnesses, workflows, and dynamic orchestration
Dynamic Workflows: the feature that enabled the Bun rewrite. Pydantic adds DynamicWorkflow so models can orchestrate swarms of agents and produce auditable, large-scale multi-agent code runs. Outcome engineers get a concrete orchestration primitive for observable, reproducible agent pipelines — Principle 09/13.
Agentic AI Foundation updates MCP with stateless architecture, hardened authentication, and 12-month deprecation policy. The MCP revamp moves to a stateless design, tightens authentication, and formalizes a 12-month deprecation window for integrations. Engineers must reassess context-layer integrations, token rotation, and lifecycle policies to avoid brittle agent deployments — Principle 06/10.
Snowflake launches Cortex AI Gateway to control AI agents and prevent runaway enterprise costs. Snowflake releases a centralized gateway that enforces agent identity, data access policies, and spend controls at the platform edge. If you operate fleets of enterprise agents, this is a pattern for enforcing data governance, identity, and cost ceilings without changing every agent — Principle 09/10.
Visa used Mythos to hunt for bugs in its own payment network, then open-sourced the harness that made it possible. Visa runs Anthropic’s Mythos for automated vulnerability hunts and publishes the agentic harness that governed test scope, artifacts, and evaluation. Steal these harness patterns to run adversarial, auditable security hunts safely and retain verifiable artifacts — Principle 07/14.
GM redesigned its engineering workflows around AI agents — and tripled its merged pull requests. GM retools developer processes so agents handle routine tasks and humans focus on higher-leverage work, dramatically increasing merged PR throughput and reducing defects. Outcome engineers should treat organizational workflow changes, CI/agent loops, and accountability artifacts as first‑class system components when scaling agentic teams — Principle 03/09.