Skip to content

@maelcaldas says they played the whole agent playbook for 18 months: great for building software, still not for running operations. After TypeSafe released Jev on 15 September, their apps were already full of Jev calls, so he put Decisions behind a standard interface on their AI gateway, with a TypeSafe-compatible spec, so they can test models without changing clients.

Apps

Mael Caldas @maelcaldas

EN

  • 40 views
  • 1 likes
  • 0 bookmarks
  • 0 reposts
We played the whole agent playbook for 18 months. Great for building software, still not for running operations. Then @typesafeai released Jev, and the light at…

About this case

Product. Mael Caldas (@maelcaldas) links his article, “Agents build our software. They still can't run our operations.” He writes that an LLM predicts the next token, agents run that in a loop, and each step multiplies error, which does not fit operations such as finance and accounting. On 15 September TypeSafe released Jev. Diogo Almeida, its founder, described a model trained for judgments: one call returns several judgments, each with a confidence, in a small fraction of the time and cost. Caldas says it was faster, cheaper, more precise and more predictable than the LLM structured outputs they had used. Their apps were already full of Jev calls, so testing another model meant changing every call. He added Decisions to their AI gateway behind a standard interface, wrapped the TypeSafe TypeScript client in a TypeSafe-compatible spec, and added adapters, on the same idea as their Open Responses spec. He later moved that spec to its own org and published it as Open System One, and says they refactored the gateway and primitives around System One models. He also names Cloudflare's Clef, OpenAI's Decisions API on top of Luna, Cloudflare's Clairic (extraction, not decisions), and Google's Gemini Decision 1. No tweet media.

Original posthttps://x.com/maelcaldas/status/2121042288315134246

Related cases