A run workbench, mission control and six new agent components
Steer an agent while it works, see its plan against what happened, watch its memory fill, and oversee a whole team of agents with spend and approvals in one place.
New
agent-11is a run workbench: the conversation with each tool call as an openable card, a live plan, a context meter and a steering box, with Stop and honest failures.agent-12is mission control for a team of agents: spend, tokens, running agents, the delegation tree and an inbox of approvals.agent-steeringredirects an agent mid-run (now, after this step or at the end) and shows each instruction going from queued to applied.plan-trackershows a plan against what happened, including steps the agent added or skipped.context-windowshows what fills an agent's memory and warns before it overflows.agent-treeshows who delegated to whom, with rolled-up progress and cost.model-comparesets models side by side by intelligence, speed and cost.thinking-loaderis a small working pill for AI replies.
Improved
tool-call-cardgets a status strip, a server tag, an icon tile, keys and values for flat inputs, Copy, a moving line while running, a Denied status and Retry on a failed call.autonomy-dial,trace-waterfall,agent-lanes,agent-graph,parallel-attempts,tool-catalogandescalation-cardget richer headers, costs, models, icons, risk badges and meters. Every component page now has a Copy prompt button for your coding agent.