Jev Engineering: turning an agent stack into a control system

Straight from TypeSafe's launch material, not from anything Ricker measured. The post labels them as being from tests, and the linked article is the vendor's own 10-step guide, which also claims that multi-step decisions taking 5 to 20 seconds become millisecond routing. No baseline task, model pair, or cost table is shown anywhere, so treat the multipliers as a vendor claim.

The post only names them. The chain runs request, structured state, Jev router, cheapest capable model, worker, relevance and approval checks, execution gate, tool, and it says the decision layers route, score, block and approve instead of writing prose. Nothing in the video defines which questions feed those two stages, so you would be inventing the policy yourself.

Just an explainer. It is a video post that links to a written article, with no repository, no code, and no package to pull. For the routing primitive itself you need the Jev docs and SDK directly. Treat this as architecture framing rather than something you can run.

Because the design does not remove generation, it demotes it. Cheap decisions route to the cheapest capable model and hand off to a worker, so the expensive model is only called when the task needs it. The post gives no data on how often that escalation fires or what accuracy you lose at the cheap tier, which means the saving depends entirely on your own thresholds.

Ricker, who describes himself as an AI researcher, developer and engineer, and whose post teases an article titled Jev Engineering: The 10-Step Guide to Giving Your Agents a Decision Brain. The diagram and multipliers restate that article, so the post is an entry point rather than an independent source.

It means keeping generative reasoning in the model that needs it while pulling the cheap yes-or-no calls into a layer that routes, scores, blocks and approves without writing prose. The stated goal is to make that decision layer measurable, cheap and fast, so the expensive model is invoked only when a task needs generation.

The router picks the path and the worker does the task. In the post's chain a request becomes structured state, a Jev router selects the cheapest capable model, that model acts as the worker, and only then do the relevance and approval checks and the execution gate decide whether the tool call runs. Routing is a decision; the worker is where generation happens.