Nova ADK

Most teams building agents rewrite the same scaffolding every time. Routing a request to the right handler, keeping conversation state, exposing tools safely, deciding what an agent may do unsupervised, and working out afterwards why it said what it said.

Nova ADK is that scaffolding as a platform. An agent is a configuration rather than a codebase, and the parts that are genuinely hard to get right, guardrails, trust evaluation and audit, are built once and shared by every agent on it.

Trust evaluation runs on output, not on faith

Every response is scored before it leaves the pipeline. An agent that is usually right and cannot tell you when it is wrong is not deployable into anything that matters, so the evaluation stage is part of the run rather than a monitoring afterthought.

Guardrails sit before the model, not after

PII detection happens on the way in. Redacting a response after the fact means the sensitive data already reached the provider, which is the part that actually matters for compliance.

Tools are typed and closed

An agent reaches the outside world only through declared tools with validated arguments. It chooses the operation; it never chooses the endpoint. Letting model output determine a URL hands a language model the ability to make arbitrary requests from inside your network.

Templates because agents repeat

A marketplace of prebuilt agents exists because most real deployments are variations on a handful of shapes. Building the base once and configuring per use case is the difference between a two week delivery and a two day one.