Monday, November 16, 2026 Edition No. 240
Search
Daily News Fully automated · 4 min read

Architecture AI Updates: November 13, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. Martin Fowler Describes the Generated Glossary: an LLM-Built, Human-Reviewed Vocabulary That Other Agents Consult

Martin Fowler published a bliki entry on November 12 describing a technique for domains whose vocabulary is absent from a model’s training data: have an LLM read an organization’s documents, emails and chat logs, identify unfamiliar terms and familiar terms used in unfamiliar ways, and draft dictionary-style definitions that domain experts then correct. The glossary is stored where other LLMs can consult it at low token cost, kept in version control, and refreshed on a schedule by agents and humans; for large glossaries, he suggests placing only the terms and links in the initial context so the model fetches full definitions on demand. Fowler advises against a single standard glossary and recommends one per Domain-Driven Design bounded context, including one derived from source code, with comparisons between glossaries used to expose differences in usage. He also suggests listing terms that should not be used, with replacements, and having harness sensors check that generated code and user interfaces follow the vocabulary. Source

2. JioHotstar Benchmark Finds No Single Code Retrieval Approach Wins for Coding Agents, and Plain Filesystem Tools Stay Competitive

InfoQ reported on November 11 on an evaluation by Abhishek Bansal, Senior Staff Engineer at JioHotstar, of six ways to help coding agents navigate large codebases: filesystem tools, Sourcebot, Graphify, CocoIndex Code, Repowise and an in-house RAG pipeline. The tests used subsets of SWE-bench and CodeScaleBench plus a custom question-answering harness over JioHotstar’s own code, with Claude Sonnet as the solving model, and measured accuracy, token consumption, agent turns and tool usability. Structural search (Sourcebot’s indexed text search, symbol definitions and references) beat the filesystem baseline on harder SWE-bench Verified tasks at lower cost, while semantic retrieval helped more on CodeScaleBench tasks with ambiguous problem descriptions; on simpler SWE-bench Lite tasks, retrieval mostly added overhead. The team recommends keeping filesystem tools as the baseline, choosing structural or semantic indexing by task type, and measuring tool-use traces alongside accuracy. A commenter quoted by InfoQ noted that the results come from limited subsets without repeated runs of the same setup. Source · JioHotstar post

3. QCon AI Talk Shows Sparse Autoencoders on the Decision Token as a Way to Explain and Monitor Agent Tool Selection

InfoQ published a QCon AI presentation on November 12 in which Hannes Hapke, who leads Dataiku’s open-source 575 Lab, describes Kiji Inspector, an open-source framework for explaining why an agent picked a given tool. Instead of relying on chain-of-thought text, it trains a sparse autoencoder on the layer activations at the single token where the model names its tool, using about one million prompts built as contrastive pairs (similar requests that should route to different tools), then validates the learned features with causal ablation. Because standard inference servers do not return activations, the project ships a patch for vLLM and containers for production use. Hapke lists prompt validation and drift monitoring (watching whether activation patterns shift over time) as production uses. Stated limits: it requires open-weight models, training is compute-intensive, and the current work covers single layers only. Source