ALGAL
Build durable agent programs with typed decision cells, replayable effects, and explicit model-routing policy.
Added to Jevfast
How it uses Jev
ALGAL can send a task prompt and typed questions through its optional Jev adapter. Its model-router example asks for difficulty, scope, ambiguity, stakes, task kind, and frontier-worthiness. A fitted expression combines the answers with text features to choose frontier or standard; probe tasks route to standard. Other providers are also supported.
What Jev decides
Illustrative model-router requestExample answers · not a recorded Jev response · Source ↗
Question 1 · frontier
YOUR APP
INSTRUCTION
Would a frontier model materially improve the result for this task?
STATE
A first-session task prompt. This source example defines the request fields and questions but does not provide a concrete runtime prompt.
JEV · NOUL
YesNo
App workflow
- Receive a task prompt
The router takes the first-session task text as input.
- Derive features and ask Jev
An expression computes prompt-shape features, then one bounded decision call asks typed questions about difficulty, scope, ambiguity, stakes, kind, and frontier-worthiness.
- Combine and route
A fitted expression combines the features and answers. A probe-kind gate selects standard; otherwise the score selects frontier or standard.
Why it is interesting
It separates fast model judgments from deterministic routing policy: a fitted expression combines the scores and a request-kind gate can override the resulting route.