Architecture AI Updates: November 18, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Shopify’s Remote Dev Gives Each Coding Agent Task Its Own Isolated Environment, With iOS and Android Builds Reached Over SSH
Shopify built Remote Dev, a remote mobile development environment where AI agents build, test and maintain iOS and Android apps, after running three or four parallel agent workflows was enough to max out a developer laptop. According to InfoQ, each task gets its own isolated environment, a remote macOS fleet runs the iOS toolchain and simulators, and Android emulators run on separate Linux VMs, all reachable through one SSH session. Agents can build, run tests, drive simulators, capture screenshots and open pull requests, and Shopify also runs overnight workflows that find bugs and propose fixes. Shopify reports that River, its Linux-based environment for web and backend work, grew from 6% to 61% of merged pull requests. Source
2. LinkedIn Scales an LLM Ranker Under a 500 ms P99 Budget by Distilling a 7B Model to 400M and Compressing Job Descriptions to One Embedding
A LinkedIn engineer presented how an LLM-based ranking stage fits into a three-step pipeline of query understanding, vector retrieval and ranking, with the ranker scoring each candidate from the probability of a “yes” token. The team distilled a 7B teacher into a roughly 400M student and cut job description input from about 2,000 tokens to 150 through offline summarization, then to a single embedding. The speaker reports throughput per GPU rising from about 290 items per second to about 2,200 with summaries and about 22,000 with embeddings, with under 2% quality loss at the summary step. Serving fixes included batch tokenization, an item-aware attention mask for scoring many candidates in one pass, and gc.freeze to remove garbage-collection pauses of about 300 ms. Source
3. Thoughtworks Proposes Deciding AI Sovereignty Per Workload Across Context, Model and Infrastructure Layers
Thoughtworks published a framework that treats sovereignty as a per-workload decision rather than an organization-wide policy. It separates a context layer (prompts, evaluations, agent skills, guardrails and policies, managed as an enterprise harness), a model layer (proprietary and open-weight models behind an agentic gateway that can route between them) and an infrastructure layer, with governance across all three. Workloads are classified as “must control,” “should control” or “can rent,” and the article suggests owning infrastructure last, using platforms such as KServe, KAITO and llm-d where they add measurable value. It also recommends mapping AI dependencies and testing at least one fallback before it is needed. Source