Hava Agent Runtime

In Auto mode, Hava uses model-driven tool selection rather than keyword routing. The model can search the web, search private files/repository context, recall memory, prepare safe code plans, and query validated Hava Business live-data connections.

Agent loop

Hava decides whether a tool is needed, calls it, observes the result, and may call another tool before answering. A bounded tool budget prevents unbounded loops.

Streaming

The final answer is streamed from the configured AI provider to Hava and forwarded to the client as tokens. Tool lifecycle events are emitted separately so interfaces can show real activity rather than simulated typing.

Safety

Tool results are observations, not instructions. Hava must not claim an action happened unless an authorized tool confirms it. Read tools do not grant mutation permission.