ARSI
How an agent gets a measured update, and why a weaker score does not publish.
The round
ARSI proposes a reply, a plan, and a tool call. A hosted judge scores them when a provider key is set; otherwise a mock outcome model does. The winner is checked against the agent's suite. Passing changes become proposals: instruction text, a skill, an example, and a memory line on the agent. Client memory notes are not rewritten.
Approval
New agents start in suggest mode. An editor approves a proposal to publish an immutable version. Auto-apply publishes without that click and still stops when the suite score would fall. Off runs nothing. Rollback restores the snapshot from before that publish, and only for the version that is current.
Workspaces
A linked workspace with no pin follows the published version. A pin stays on the version it names. The dashboard charts suite scores by round.