The Model Note https://themodelnote.com/ Practical notes for choosing and deploying AI models and agents — transparent decision guides, workflow tools, and implementation write-ups. en Stop asking agents for permission at every step. Start deciding where approval belongs. https://themodelnote.com/news/where-approval-belongs/ https://themodelnote.com/news/where-approval-belongs/ Sat, 22 Aug 2026 00:00:00 GMT Three approval models—per action, per plan, and per boundary—and a practical default for a small team running its first supervised agent pilot. Before you add another agent, build a five-case evaluation suite. https://themodelnote.com/news/a-small-eval-suite/ https://themodelnote.com/news/a-small-eval-suite/ Sat, 22 Aug 2026 00:00:00 GMT A compact way to move beyond glossy demos: real tasks, observable outcomes, trace review, a failure case, and repeatability. The most useful first agent may be a workflow that does not look autonomous. https://themodelnote.com/news/workflow-before-agent/ https://themodelnote.com/news/workflow-before-agent/ Sat, 22 Aug 2026 00:00:00 GMT A decision guide for choosing the least-complex AI system that can improve a recurring task—and knowing when an agent is actually justified. Before you trust an agent, make its work easy to replay. https://themodelnote.com/news/trace-before-trust/ https://themodelnote.com/news/trace-before-trust/ Sat, 22 Aug 2026 00:00:00 GMT A practical review packet helps a team inspect the request, boundary, path, outcome, and human decision—without mistaking a polished summary for evidence. Context is a budget. Spend it on the next decision. https://themodelnote.com/news/context-is-a-budget/ https://themodelnote.com/news/context-is-a-budget/ Sat, 22 Aug 2026 00:00:00 GMT A practical way to give an agent the smallest trustworthy packet for the job, then leave a handoff a human or fresh session can actually use. We spent $0.05 testing Jev. Its calibration held. The pipeline around it didn’t. https://themodelnote.com/news/jev-calibration-test/ https://themodelnote.com/news/jev-calibration-test/ Tue, 22 Sep 2026 00:00:00 GMT A pre-registered 1,000-call check of TypeSafe’s System One model: stated confidence tracked accuracy in every bin we could test, and it answered all 1,150 calls with typed output — while a chat baseline failed to return usable JSON on a third of calls. We gave an AI agent full disk access. Here’s what it could read. https://themodelnote.com/news/os-agent-guardrail-fda/ https://themodelnote.com/news/os-agent-guardrail-fda/ Mon, 05 Oct 2026 00:00:00 GMT After Apple pledged stronger Full Disk Access consent over AI-agent abuse, we ran a data-access census with an autonomous agent on a stock Mac: 3,554 message rows, browsing history and notes read silently, a file written into a protected directory, and the system permission table enumerated — all without one prompt.