Jev vs function calling

Function calling — OpenAI tools, Gemini function declarations, Anthropic tool use — teaches a generator to emit a structured call instead of prose. It is a real improvement over “please reply with JSON.” It is still a language model choosing tokens, often with a grammar overlay. Vercel, LangChain, and Microsoft loop docs exist because the generator will keep calling. Jev vs function calling is whether the next verb is a sampled tool name or a Choice over an allowlist.

Failure modes you already have in production: a tool name that is not in the list, arguments that fail your schema, a call that should have been “stop,” a call that invents a file path. You write retry loops and parsers around that. Strict structured outputs on tools reduced the JSON tax. They did not give you peakedness, and they did not make stop a first-class option unless you invent a dummy tool. Jev vs function calling keeps that dummy as a real key: stop.

Feature contrast

What Jev changes

A Choice over search | retrieve | write_file | run_tests | ask_user | stop cannot emit run_shell. The answer space is the criteria object. You still execute the tool in your runtime. You still have to fill arguments — Jev will not write them. What you get is the decision, with a distribution you can threshold. That is Jev vs function calling in production terms.

if (next_tool.confidence < 0.70) {
  return askUser(next_tool.probabilities);
}
return execute(next_tool.choice);

A two-step agent (Jev vs function calling in practice)

Step A: Jev Choice over next tools, including stop and ask_user. If confidence is low, do not execute. Step B: if the Choice is a tool that needs arguments, call a writer or a dedicated extractor to fill the schema, then validate with your usual JSON Schema / Zod. Execute in your runtime. Then Jev again for continue or stop. That is slower to draw on a whiteboard than “one tool-calling loop,” and it is how you stop paying for four identical searches.

Keep function calling if a single model turn must both talk to a user and pick a tool — product chat, copilot sidebar. Jev vs function calling is the wrong fight in that UI. Use Jev when the agent is a machine loop nobody is watching, where an off-list tool is an incident and stop must be cheap.

See the agent tool selection pattern and continue or stop. Related: JSON mode, ChatGPT. Independent — not affiliated with TypeSafe AI. Last reviewed 17 September 2026.

Loop control: Jev vs function calling on stop

Most agent frameworks stop on max_iterations, a finish reason, or a dummy tool with no execute. Those are runtime brakes. They fire late. Jev vs function calling on stop is asking a Choice after each tool result: continue, retry, ask_user, stop. You still keep max_steps = 8. You now have a peakedness number three steps earlier than the cap. That is the difference between a budget incident and a planned handoff.

Off-list tools are the other incident. A generator with a long tool list will eventually emit a cousin name. Strict schemas help. An allowlist Choice cannot emit the cousin. If your security story is “the model will not call the shell,” put the shell off the list and use Jev for the pick. If your security story is “the model may call the shell but arguments are reviewed,” that is still function calling, with humans on arguments — Jev vs function calling does not replace that review.

Try agent tool selection in Jev Studio and when not to use Jev so you do not ask the Choice to write the patch.