AI Architecture Updates: May 22, 2026

1. Martin Fowler draws an architectural line between vibe coding and agentic programming

Martin Fowler. A new bliki entry treats vibe coding, the practice of prompting an LLM to build software without reading the generated code, as architecturally distinct from agentic programming, where developers stay in the review loop. Fowler argues the approach is acceptable for disposable applications with trusted users, but warns that on consequential systems it accumulates poorly structured code that becomes difficult to modify, exposes secrets and data through what he calls the “lethal trifecta,” and amplifies correctness risk because LLM non-determinism turns every enhancement into a potential regression. The framing positions code review and active maintenance as non-negotiable architectural controls, not optional polish, whenever software has to evolve or carry sensitive responsibilities. Source

2. Grab case study on InfoQ shows supervisor topology and tool consolidation as the load-bearing choices for multi-agent support

Leela Kumili on InfoQ. Coverage of Grab’s engineering-support platform details a supervisor-based multi-agent design built on LangGraph and FastAPI that splits incoming work into two paths: an investigation flow for diagnostics and an enhancement flow for code changes. The architecture deliberately constrains each specialized agent to a narrow responsibility, compresses context with structured strategies to stay within token budgets, and consolidates around 30 tools down to a smaller curated subset because broad tool menus made agent selection unpredictable. Safety layers include SQL validation, sensitive-data detection, and mandatory human review before any code lands, and Grab credits the system with reclaiming hundreds of engineering hours per month. The takeaway is that scoping, tool curation, and human checkpoints matter more than model choice when moving multi-agent systems from demo to production. Source

3. Meryem Arik makes the case for the AI gateway as a first-class architectural seam

Meryem Arik on InfoQ. A new presentation frames the “AI gateway” as the architectural seam organizations need to tame inference chaos as decentralized teams adopt many models in parallel. The pattern centralizes cross-cutting concerns including RBAC, cost attribution, key management, and traffic shaping into a thin layer in front of model providers, leaving teams free to pick the best model per use case without each one rebuilding governance. Arik points to open-source implementations like LiteLLM and Doubleword as concrete starting points, and positions the gateway alongside service meshes and API gateways as a standard piece of platform infrastructure rather than a bespoke wrapper. Source