Stop paying frontier prices for every line of codeJev routing picks the right model for each task
AutoZorg JEV uses Jev, a decision model from TypeSafe, to route each coding task to the least costly tier that can still do the job well — before any tokens are spent on generation.
The hidden tax in AI coding agents
An AI coding agent does not answer once. It reads files, runs commands, edits code, checks the result and repeats — and a single request can trigger dozens of model calls.
Every call re-sends a growing conversation as input tokens, so your choice of model is multiplied across the whole loop. If a frontier model handles a task a mid-tier model could finish, you are not overpaying once — you are overpaying on every step of that loop.
Routing fixes the problem at the source: decide which model should do the work before the loop starts.
What is Jev?
Jev is a decision model from TypeSafe. Unlike a chat model, it does not write text — it returns typed decisions built for the job of choosing.
Typed decisions, not prose
Jev returns a structured choice rather than paragraphs of reasoning you would pay for. It is a classifier built for choosing, not a generator repurposed for it.
A probability distribution
Every decision comes with a distribution across the six coding tiers, so you can see how the options compared, not just the winner.
A confidence value
Jev reports how concentrated its output is. If confidence falls below a threshold, the run pauses and asks for clarification instead of guessing.
What is Jev routing?
Jev routing places Jev in front of a group of coding models and lets it pick which one should handle a given task. In AutoZorg JEV, the options are six coding tiers, ordered from least to most costly, and Jev selects the least costly tier that can still do the job well.
A simple copy change or a styling tweak can land on a lower tier. A multi-file refactor or a tricky bug can go to a stronger, more expensive model. The routing decision is made once per user turn, and then the selected model takes over.
How AutoZorg JEV works inside ZorgSocial Builder
Here is what happens when you send a message with AutoZorg JEV selected.
Jev routes the task
Jev reads the task and the shared conversation, then chooses one of six coding tiers before any generation happens.
The coding model takes over
The selected model works through its normal tool loop with the same sandbox tools: reading, writing and editing files, searching the codebase and running shell commands.
The turn is saved
The coding model finishes with its own final response, which is stored in your session transcript.
Jev routes again
On your next message Jev decides again. If the task is harder or easier, a different tier can pick it up.
Two safeguards keep this predictable. If Jev's confidence falls below a set threshold, the run pauses and asks for clarification instead of guessing. If the router is unavailable or returns an invalid answer, AutoZorg falls back to a middle tier so your work keeps moving.
Why Jev routing reduces token costs
Routine work stops paying premium prices
Most real app-building requests are small: adjust spacing, fix a typo, add a field, change a color. Routing sends these to cheaper tiers, where most of the savings come from.
The router stays out of the expensive loop
In production mode, Jev makes one decision per turn. It does not sit in front of every tool call, so routing overhead stays small and fixed instead of growing with the task.
Switching tiers keeps your context
All six tiers share one transcript, so a new model inherits the full history of messages and tool results — no re-reading files or re-explaining the project.
Decisions are cheap by design
Because Jev returns a typed choice rather than generated text, the cost of the routing step itself stays low compared with the coding work it directs.
How much you save depends on your task mix. A project full of small edits benefits far more than one made entirely of hard architectural work — so every AutoZorg run reports its routing decision, the selected tier and the Jev usage cost in the Builder's run inspector.
What Jev routing does not do
We think honesty about limits builds more trust than marketing claims, so here is the fine print.
Jev's confidence score describes how concentrated its output distribution is. It is not a guaranteed probability that the chosen tier is correct. Jev also cannot run your tests or inspect your code, so it does not prove that a change works. That remains the job of the coding model, your build, and your own review. Routing is a cost and efficiency tool, not a replacement for testing.
Routing-only mode vs guarded mode
Routing-only mode (default)
Jev picks the tier and the coding model works freely from there. This keeps the builder fast and lets the coding model recover from normal errors on its own.
Guarded mode
A stricter mode for testing and high-stakes workflows. Jev checks each tool call before it runs and verifies each result afterwards, holding anything unsafe or out of scope. It is not the default, because gating every step adds latency and cost.
Frequently asked questions
What is Jev in AI?
Jev is a decision model from TypeSafe. It returns typed decisions, such as a choice among options with a probability distribution, rather than generating text.
How does Jev routing save money?
It sends each task to the least costly model that can handle it, so routine requests avoid premium model pricing. Because a coding agent makes many model calls per request, choosing a cheaper model up front reduces spend across the whole loop.
Does switching between models lose my project context?
Not in AutoZorg JEV. All six tiers share one conversation transcript, so a new tier sees the previous messages and tool results.
What happens if Jev is unsure?
If confidence falls below the threshold, the run stops and asks for clarification. If Jev is unavailable, AutoZorg falls back to a middle tier.
Can I see which model handled my request?
Yes. The Builder's run inspector shows the tier chosen, the probability distribution, the selected coding agent and the Jev usage cost.
Build your next app with Jev routing
Describe an app in plain language, watch a full Next.js and TypeScript project come to life in a live preview, and deploy it to your own domain with SSL. Choose AutoZorg JEV and let Jev decide which model each request needs.