An agent that acts on enterprise data has to be reviewable while it works, not only after it stops. Reading a transcript once the run is over tells an operator what happened; it does not let them intervene. This release makes the run itself observable: the lifecycle of every turn is published to observers as it happens, the model's reasoning is streamed rather than withheld until the answer, and the plan the agent commits to is a durable artifact rather than a line in a log.
A live turn lifecycle
- An observation point on the run. A run publishes its turn lifecycle to registered observers, so a client, an audit sink and a persistence layer each see the same events without any of them owning the run.
- One path for streaming and observation. A component that both streams tokens and observes lifecycle events is adopted as an observer automatically, so a streaming client does not have to register twice to see what the run is doing.
- Server-sent events on the same model. The browser's event stream and the conversation's stored history are both driven from this lifecycle, so what a reviewer watches live and what is written to the record cannot drift apart.
Reasoning that arrives as it is produced
- Streamed thinking. Where a provider exposes its reasoning content, that content is streamed to the client as it is generated, delivered on a dedicated
THINKING_CHUNKevent so it is rendered apart from the answer. - Visible before the outcome. A reviewer sees the shape of the model's approach while the run is still in progress, which is the point at which stopping it costs nothing.
The plan as a stored artifact
- Written down, then ticked off. An agent's plan is persisted as a file it updates as it works, so progress against the intended sequence is inspectable rather than inferred from what has run so far.
- Typed and validated. The plan carries a defined schema and is validated on write, so a malformed or partial plan is rejected instead of stored.
- Classified in the file store. Stored files carry a kind and a visibility, so a plan is distinguished from an ordinary attachment: it stays out of the general document library and is downloadable only by the people granted access to that conversation.
The change is additive for existing conversations: an agent run without a registered observer behaves exactly as before, and reasoning content is streamed only where the configured provider makes it available. What is new is that supervision no longer has to be reconstructed after the fact.