An AI agent that can act on a database is only safe when its actions are governed the way any other privileged operation is. This release adds three controls to the Studio agent: it can propose a plan before it acts, it runs under a chosen model per conversation, and every tool call it makes is logged as it happens.
Plan mode
- Propose before acting. In plan mode the agent drafts the steps it intends to take and records them to a scratchpad first, so a reviewer sees the intended sequence before any step runs.
- An explicit gate. The plan is a checkpoint, not a formality: the operator reads the proposed steps and decides whether the agent proceeds.
Per-conversation model selection
- Choose the model per task. Each conversation carries its own model and mode, requested at the point the conversation starts and drawn from the list of available models.
- Bounded iterations. The agent's maximum iteration count is configurable, so a run carries a ceiling rather than looping until it decides to stop.
Real-time tool-call logging
- Logged on the fly. Each tool the agent invokes is written to the log as the call happens, rather than only after the run completes.
- Pollable stream. The
stream_startevent carries the message id, which a client uses to poll the tool-call log for that message in real time.
Together the three controls make the agent a supervised actor: its plan is reviewable, its model is chosen, its iterations are bounded, and its actions are visible while they run.