Architecture AI Updates: October 1, 2026

1. An Agent Gateway Pattern for Governing MCP Tool Access

ByteByteGo. A breakdown of DoorDash’s Agent Gateway argues that MCP standardizes tool discovery and invocation but leaves permissions, credentials, and monitoring to each team. DoorDash splits the problem into access, tool-surface curation, and operations, with a data-plane proxy that authenticates callers, applies rate limits, attaches credentials, and logs calls, and a control-plane registry that holds agents, servers, policies, and curated tool catalogs. Separate proxy planes for internal and external use keep trust boundaries apart while sharing the same governance model. Source

2. A Replay Harness Made a 300K-Line Agentic Refactor Verifiable, and Hard to Generalize

InfoQ. InfoQ reports on CodeScene’s case study in which coding agents refactored the 300,000-line C decompilation of Street Fighter III: 3rd Strike over three weeks for about $4,000 in tokens, with 2,903 commits and a Code Health score rising from 5.6 to 10.0. Each change was checked by a replay-trace harness that compared rollback state hashes frame by frame, and the agents accumulated a playbook of 22 refactoring recipes. Practitioners noted that deterministic, frame-level verification is rarely available in legacy business systems and that architectural health and performance were not measured. Source

3. Treating Agent Context Like Code

InfoQ. In a QCon talk published by InfoQ, Patrick Debois makes the case for managing coding agent context with the same practices used for code: tests, CI/CD, package managers, and security scanning. The argument is that versioned, evaluated, and observable context is how engineering organizations scale agent use while keeping non-deterministic output under control and building up reusable organizational knowledge. Source