Agent workspace
uses 2 of today's live runsA bounded agent: a planner picks tools (document search, a safe calculator, read-only SQL over a catalog of the knowledge base), the tools run in parallel, a writer answers only from the tools' evidence with citations, and the evaluator verifies the answer, allowing one correction. Hard limits apply to steps, LLM calls and time, and retrieved text is treated as data, never as instructions.
Search