Architecture AI Updates: November 20, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Thoughtworks Describes an “AI Factory Stack” and Argues Token Economics Is Infrastructure Economics
Thoughtworks published an essay by Vanitha Kumar arguing that the per-token price of a model is one line item, and that routing, batching, storage, network, scheduling, isolation and failure handling decide the rest of an AI bill. It defines the AI factory as the production system that turns compute, data, storage, networking and models into priced inference, with four jobs: utilization, latency, resilience and attribution. The essay maps the stack to layers including routing and policy (LiteLLM, agentgateway), serving and scheduling (KServe, KAITO, llm-d), observability (VictoriaMetrics), isolation (vCluster, agent sandboxing) and workflows (Temporal), citing their ratings in volume 35 of the Thoughtworks Technology Radar. It notes that agent loops re-read context on every pass, so cache-aware routing and gateway policy matter more at agent volume than at chat volume. Source
2. InfoQ Reports a GOTO Talk Arguing Healthy Code Lets AI Agents Work With 40% to 45% Fewer Tokens
InfoQ summarized Adam Tornhill’s GOTO Copenhagen talk on AI-friendly code. Tornhill cited a study of more than 1,000 open-source projects in which AI-driven velocity gains in the first weeks were fully canceled out after about two months by rising code complexity. He said the average codebase scores 5.15 on a code health scale while AI-friendly code starts at 9.5, that an agent alone fixes roughly 50% of code health issues in lower-health code and 20% once health improves, and that adding an MCP server that supplies code health data raised the fix rate to 90% to 100%. He also said healthy code needs 40% to 45% fewer tokens. The accompanying interview describes safeguards exposed to agents as local MCP servers or command-line tools, and a refactoring loop of baseline, behavioral safety net and automated feedback. Source
3. Pinterest Staff Engineer Describes Scheduled Sandboxed Agents That Run Code Migrations Across Teams
InfoQ covered a QCon San Francisco 2026 talk by Jordan Cutler of Pinterest on how a staff engineer’s leverage changes when every engineer has AI tools. He presented internal data showing merged pull requests per engineer per week rose 68% for the top 10% of engineers (6.7 to 11.3) and 14% for the bottom half between Q1 2025 and Q3 2026. His team built a system where a staff engineer supplies a plan and an example, an engineer schedules it, and an agent in a sandbox splits the migration into pull requests team by team; a class component migration ran three to five times faster than by hand, and 40 engineers now run their own scheduled jobs. One looping job opened 16 pull requests in two minutes and raised CI wait from under 30 seconds to 14 minutes, which the speaker used to argue for shipping a small guardrail early. Source