Reading note. Our take on published research, with sources below.
If you run a lean ops or support team, the Jev wave is less about a new brand and more about a shape of work you already have: many small decisions, each with a short list of allowed outcomes. Here is a practical map of where a System One–style model belongs — and where it does not.
Good fits
Customer message → which queue. Refund request → auto-approve, deny, or human review. Monitoring alert → page now, batch, or ignore. Agent step → which tool among a fixed list. Browser or UI agent → which visible control to activate next.
Community demos that spread this week — ad classification batches, flight-booking agents, Claude Code harness checks — all share that shape: state in, closed set of answers out, often under a few hundred milliseconds per call according to public write-ups.
Weak fits
Anything that must produce customer-facing language, long reasoning traces, or novel plans still wants a chat or reasoning model. Jev’s own pitch gives up string generation. Forcing a decision model to “write the reply” is the wrong category.
Also weak: one-off decisions with no volume. If you make a judgment twice a week, a human or a simple rule may already be cheaper than integrating a new model.
Compose with code, keep a person for edges
TypeSafe’s workflow evals argue for decomposing tasks into narrow questions and combining results in code. That matches how durable automation already works: models for judgment, programs for structure, humans for exceptions.
Set confidence thresholds so high-confidence paths run automatically and low-confidence paths open a ticket or Slack thread. Log the state, the question, the answer, and the action taken — otherwise you cannot improve the policy.
How we would start with a client
Inventory the top recurring decisions. Pick one with clear options and a painful current cost (time or error rate). Define questions and thresholds. Shadow-run against last month’s cases. Only then put it on the live path.