Architecture AI Updates: October 20, 2026

1. From Microservices to Agents: Separate Planning From Execution and Bound the Loops

InfoQ (Mallika Rao). An InfoQ article in its Agentic AI Architecture series argues that microservices decomposed functionality while agentic systems decompose decision-making, which moves control flow from design time to runtime. The recommended structure separates a planning component that interprets intent from deterministic components for retrieval, transformation and ranking, puts standardized tool interfaces between reasoning and services, and keeps a deterministic fallback pipeline for low-confidence cases. Evaluation loops should be bounded with latency budgets, iteration limits and confidence thresholds to avoid retry amplification. The author also calls for decision-level observability (tool selections, retry rates, fallback rates, strategy-selection frequencies), since silent quality degradation can occur while latency and error metrics stay healthy. Source

2. Inside Uber’s Genie: RAG Over Internal Docs, LLM Judges for Answers and for the Docs Themselves

ByteByteGo. ByteByteGo breaks down Genie, Uber’s internal Slack support assistant, which uses RAG instead of fine-tuning so knowledge can be updated without retraining. A Spark ETL pipeline chunks Engwiki and internal Stack Overflow content, embeds it and loads it into Uber’s Sia vector database while keeping source URLs, and each retrieved passage is passed to the model as a separate sub-context that must be cited. Evaluation combines Slack feedback buttons (Resolved, Helpful, Not Helpful, Not Relevant) with an LLM-as-judge for hallucination and relevance, and a second judge that scores the source documents and suggests edits. Uber’s reported figures: 154 channels and more than 70,000 questions answered in the first year, a 48.9% helpfulness rate and about 13,000 engineering hours saved. Source

3. Local-First Small Models, Golden Datasets in CI, and Agents That Assemble UIs

InfoQ Podcast (Rachel Lee Nabors). On the InfoQ podcast, former React core team member Rachel Lee Nabors describes building Mima, a local-first social aggregator that runs a small Llama 3.2-class model on device for summarization, which matched a large cloud model on that task after evaluation. Her evaluation method is a golden dataset from real content judged by a large model, several runs per candidate to account for non-determinism, and tracking accuracy and latency in CI/CD so new models can be swapped in. She notes local-first does not guarantee lower latency on weak hardware and suggests device detection with cloud fallback. She also discusses A2UI, a proposed approach in which an API sends an assistant a JSON schema and the assistant renders the UI from its own components. Source