Field notes from real rollouts: security, spend, agents, evals, and the operating model. Search the archive, or just ask your question.
The demo always looks good. Here is the eval playbook mid-market teams use to measure agents, catch regressions, and improve them on purpose.
Ryan Drake
Jul 14, 2026 · 9 min read
Fable 5 vanished for 19 days. GPT-5.6 launched gated. The past month is the strongest case yet for multi-model, and for not over-hedging.
AI token bills are repeating the early cloud story. How to control spend without rationing your way back to manual work, and what to measure instead.
Agents now outproduce the humans checking them. Here is the framework we use to review a full day of agent output without reading all of it.
A security architecture playbook for agentic AI: data tiers, contract terms, permission modes, and the regulated-industry rollout pattern.
A two-lane operating model for governing AI: let anyone build in a sandbox, gate anything others depend on, and make every leader own the results.
A three-stage AI adoption roadmap with checkable markers per stage, and how to move a whole organisation through it without leaving anyone behind.
You do not need a data warehouse rebuild to make AI useful. You need a thin retrieval layer, a capture habit, and honest use-case prioritisation.
Every issue: one AI workflow you can copy, what changed in the model landscape, and what it means for your business. Unsubscribe any time.
Join 2,500+ operators reading the Workflow every week.