Agentic model

The runtime uses the Vercel AI SDK provider abstraction. Current provider adapters include Anthropic, OpenAI, Google, and an OpenAI-compatible endpoint.

Generative boundaries

The runtime makes a model call only inside a generative action or Intent Classification. Deterministic actions do not pass their content through a model.

Basic reply uses one chat call and no tools. Search knowledge uses a bounded tool loop.

Search knowledge loop

1. The model receives the Assistant instructions and available tool descriptions.
2. The model searches Knowledge or reads a selected Source window.
3. The runtime records tool lifecycle events.
4. The model can repeat the search within the step limit.
5. The runtime generates the final answer and named citations.

The Thinking panel represents these recorded progress and tool events. Simplified thinking rewrites progress text for the Visitor display.

Pre-flight routing

The pre-flight is an optional decision that runs once per Visitor message, before Intent Classification. It answers seven fixed questions in one call. The questions are the Flow, the FAQ, a request for a person, the help desk, the reasoning level, the frustration level, and the language.

Each answer carries a confidence value. When an answer is above its threshold, the runtime acts on it without the classification call:

• A Flow above threshold runs directly.
• A matching FAQ above threshold returns the stored FAQ answer verbatim, with the FAQ as its Source.
• A request for a person above threshold opens the escalation menu on the recommended help desk.

Below the threshold, the turn runs as it does without the pre-flight. The Thinking panel then shows one fixed sentence for the outcome in the language of the message. The sentence never contains a Flow name.

The pre-flight is off by default. The PREFLIGHT_SHADOW environment variable set to 1 records decisions without acting on them. The PREFLIGHT_ROUTING environment variable set to 1 acts on them. A nightly job replays a labelled set of messages and raises an Alert when a decision changes.

API integration tools

One Assistant can register one API catalogue. The catalogue describes its base URL, protected credential, and available endpoints.

The model can inspect API details and query a registered endpoint. The runtime validates every path against the catalogue before network egress.

Windowed readers return a requested content range and total length. This contract lets the model inspect large API and Knowledge payloads without silent truncation.

Failure behavior

The keyword router handles classification when no classifier model is available or classification fails. Knowledge search can use lexical fallback without embeddings.

An unconfigured email transport returns an honest failure result. It does not create a successful delivery message.

Extension rule

Add a new agent tool through the tool registry and its lifecycle wrapper. Keep authorization, ownership, input validation, and egress checks outside model control.