Introducing JevAgent

JevAgent is an optional, opinionated agent in the public Vidbyte Python SDK. It keeps the familiar model-and-tool loop, then lets you enable named Jev-backed capabilities around that loop for situations where a request needs a check before work starts or before the agent finishes.

You configure it with typed JevAgentSettings and, when needed, JevRuntimeSettings. The capabilities are fixed and validated by the SDK rather than being an open-ended framework for supplying your own questions and actions.

• Check whether a request is clear enough to start
• Route a task to a configured specialist or select from attached tools
• Check whether each requested deliverable appears complete, with bounded continuations

JevAgent does not replace the model you configure to write responses and use tools. That generative model remains responsible for the ordinary agent loop. Jev, through the TypeSafe decision model, handles the specific decision capabilities you enable; it does not become the agent's general-purpose response model.

That separation leaves the main agent's prompt, model, tools, permissions, and loop settings in your hands. Jev adds a few named checks around execution instead of taking ownership of the whole application.

When the clarity capability is enabled, JevAgent can check a request before the main model starts. If important details are missing, the run can return focused clarification questions instead of starting the task with an ambiguous brief.

That run ends with the questions. Your application must collect the answers and start a later run with the original request and those answers; JevAgent does not pause and resume the first run automatically.

If you register specialist agents with JevAgent, it can choose a configured specialist for a request or leave the task with the main agent. The specialist's scope is supplied by you; JevAgent does not invent a new agent for the job.

You can also enable tool selection to narrow the attached tool catalog before the main loop begins. It chooses among tools you already configured, so the agent's available capabilities remain grounded in your application's own setup.

For a multi-part request, an optional done check can look at whether the agent's work shows each requested deliverable. If a check finds a part missing, a bounded continuation can send the main agent back into the same loop to address it, up to the limit you configure.

This checks whether the requested pieces appear in the work; it does not verify factual accuracy. JevAgent makes completion easier to inspect, while your application still owns any domain-specific validation the result requires.

JevAgent is opt-in. With no Jev capability enabled, it makes no Jev call and follows the ordinary generative model loop. You do not need TypeSafe credentials merely to construct the settings or the agent; credentials matter when an enabled capability actually calls Jev.

JevAgent is part of the public Vidbyte Python SDK, not a hosted agent service or one of the SDK's example templates. Explore the source or start with the broader SDK introduction below.