Most of an agent's day is small decisions: is this request about billing or about a bug, is this output safe to send, which tool next. Writing a paragraph to answer each is slow and costly. Google's Agent Development Kit (ADK) 2.13.0, dated 22 October 2026 and published on 23 October, adds a classifier that sends such decisions to TypeSafe's Jev, a model built to pick from fixed options. Watch: the idea is sound, but the model is in early access and every performance figure comes from its maker.
What it is
The release lists a new TypesafeClassifier and ClassifierNode "to route workflows with TypeSafe's Jev models". TypeSafe describes Jev as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out". The possible outputs and their structure are defined in advance, so the model gives up free text: it answers with one of a set of options and a probability for each. The same release also promotes twelve experimental features to stable and changes the default telemetry format, which breaks anything that depends on the old call_llm span until you set an environment variable.
What TypeSafe claims
The company reports 70 to 500 milliseconds per call, against 3 to 329 seconds for the large language models it compared with, and calls Jev "40x-200x faster" on this kind of query. It prices input at $0.042 per million tokens, with free output, against $0.20 to $10 for the models. All of these are TypeSafe's own measurements, and the post lists its own caveats: speed was measured from the founders' laptops, "we can't prove it isn't subsidized", its test workflows were written by its own staff, the reference answers favour OpenAI and Anthropic models, and the claim that it never produces a type error is argued from how the schema matches, not tested.
What I could not confirm
Access is by waitlist in early access. The post does not say whether the model is open or closed, and its example playground is open only to early-access users, so I could not see a worked input and output. The post does not mention ADK; the integration comes from the ADK release notes.
What to do
Watch. Try it on one routing step, behind an interface that lets you swap in an ordinary model if the claims do not hold. Do not move guardrail checks to it on the strength of a vendor's benchmark.