Architecture AI Updates: October 29, 2026

1. Uber’s MCP Gateway: Registry as Control Plane, Proxy as Data Plane

ByteByteGo. ByteByteGo broke down the MCP Gateway Uber built after teams kept wiring their own MCP integrations across services that speak HTTP, gRPC and TChannel. The design separates an MCP Registry (control plane: the catalog of servers and tools, enablement and versioned config) from a Proxy Gateway (data plane) that refreshes its in-memory mappings periodically, so changes apply without restarts. AutoCrawler, a Cadence-based workflow, generates tool definitions from existing IDL service contracts, using an LLM to write better descriptions, and discovers native MCP servers through heartbeat metrics and listTools. Discovered tools start disabled until the owning team reviews and approves them, with every edit a reviewable diff that can be rolled back. The gateway exposes virtual MCP servers, converts JSON arguments to Protobuf or other wire formats, enforces caller-based policies down to individual tools, redacts PII, and exchanges internal tokens for third-party credentials for services such as Jira. To keep thousands of tools from filling the context window, Uber added Omni MCP for gradual discovery (discover_server, discover_tools, get_tool_schema), GraphQL-style response projection that trims result fields, and a code mode for coding agents. Source

2. Fowler on Supervisory Control as a Model for Agentic Coding

Martin Fowler. Fowler’s October 28 fragments include Annie Vella’s framing of the shift to coding agents as a move from direct control, where the developer closes the feedback loop, to supervisory control, where a computer closes it and the human monitors and takes over when needed. Vella maps Endsley and Kiris’s five automation levels onto AI-assisted development: no assistance, autocomplete-style decision support, agents whose actions must be approved, auto mode with a human veto, and full automation as in a dark software factory. She notes that this automation research assumed the system reports its own actions accurately, which AI agents do not guarantee, so they arguably need more supervision. Fowler also reports from Foocamp that coaches running one- to two-week agentic engineering immersion courses find the biggest hurdle is emotional rather than technical. He adds that large enterprises are held back by entrenched workflows, security concerns including the lethal trifecta, and data not yet in agent-ready shape. Source

3. Airbnb Moves Transformer Embedding Inference Into the Streaming Feature Pipeline

InfoQ. InfoQ covered how Airbnb extended its open-source Chronon feature platform so its JourneyFormer sequence recommender, a Transformer encoder over guests’ bookings, reviews and browsing, updates embeddings within seconds instead of once a day. Push Mode has the streaming job publish a notification when a feature value is committed, and NRT Model Transform runs the already-deployed model inside that streaming flow, so no separate inference implementation is needed. New embeddings land in the same low-latency store as the batch path, leaving the online ranking path unchanged and the Transformer outside the request path. Staleness fell from nearly two days to under a minute, typically 10 to 30 seconds, with a 1.67% gain in offline NDCG and about 0.33% more uncancelled bookings in A/B tests. InfoQ compares the approach with Feast’s cached-prediction pattern, which triggers inference on feature updates and trades extra writes and pipeline complexity for lower serving latency. Both features have been contributed to the public Chronon repository. Source