Field notes from the lab.
What we learn on real problems, written up plainly. Each note reviews the evidence, says what holds up, and lists its sources.
Organizational context: giving AI agents a working map of the company
Enterprise agents often fail for lack of context as well as capability. We review evidence from 2025–2026 on knowledge graphs, semantic layers, permissions, freshness and MCP, and set out what to build and what to measure.
System One decision models: Jev and the case for classifiers
Many decisions inside an AI system are choices, not essays. We review TypeSafe's Jev, the first third-party tests and the classifier toolkit that came before it, and explain how decision models with thresholds measured on your own data can cut cost and make agents easier to audit.
Small language models for AI agents: what the early 2026 evidence suggests
Many calls an AI agent makes are short, structured and routine. Early studies, mostly preprints, suggest small open models can handle much of that work at much lower cost than flagship models, but they still struggle with long, multi-step tasks. Here is the evidence, and how to test it on your own traffic.
Retrieval is not memory: what recent research shows about agent memory
Agents that only search documents do not learn from their own runs. We review recent research on agent memory: how memories are written, updated and forgotten, what the benchmarks really show, the new risks, and what to do in production.