We built a new harness using @typesafeai's Jev that cuts the cost of repetitive work by 90%. The harness learns the job as it runs, moving steps from LLM calls to code. Running 100,000 compliance alerts costs >$290K on Opus 5. With agentrun() we got it down to <$26K.

We just released an open-source version of the Jev-powered workflow DSL used in our harness. You can use it to turn your own agent into a workflow! Check it out here: agentrun.ai/

Most of the AI we use is built for a person to talk to. What changes when software is the user? In a podcast this week, Diogo Almeida explains why Jev, the new System One model from TypeSafe AI is designed for code to consume its output. But what does that mean? I keep coming back to the spreadsheet analogy. Think of each Jev call as a cell: a small, cheap decision whose result can feed into other cells. A spreadsheet cell does very little on its own. Wire thousands together and you can model an entire business. Now imagine those cells can make judgments. Does this document contain the information we need? Do these two records refer to the same company? Which queue should this case go into? Code combines the answers into a larger workflow. What makes Jev so good at this type of work? It uses post-training process called RLCD which stands for Reinforcement Learning for Calibrated Decisions. The aim is to train a model to make useful decisions and represent its uncertainty honestly. If it assigns an outcome an 80% probability across comparable cases, that outcome should happen roughly 80% of the time. That gives software something to work with: thresholds for acting, gathering more evidence, or asking a person to review. You still need to test those thresholds on your own cases. Diogo's focus on “intelligence per dollar” feels especially practical here. Frontier benchmark scores tell us something about capability. When I'm building a workflow, I also want to know how much useful work we can do at a cost and error rate we can live with. There are so many small decisions we never automate because the economics don't make sense. Make those decisions cheap enough, calibrated enough, and easy to combine, and I think we could see a Cambrian explosion of new automation. Whole workflows assembled from steps that barely seemed worth building individually. That's the possibility I'm excited about. That's we're using Jev in the new AgentRun harness on GREP.AI (YC F26). By using it to route decisions in a workflow driven harness we can cut token consumption by up to 90% while improving accuracy! Jev is available in Grep's agent harness, too.