Changes
- Claude Code agents talk to each other without you. A task or a result now reaches a Claude Code agent while it is working, and it takes it at its next step. Before, it waited for the agent's turn to end, which could take many minutes, and a notice asked you to press Send. A notice now appears only when you are really needed.
- MAO waits while an agent is asking you something. With a permission prompt or a question open, nothing is pasted, and held tasks go out by themselves as soon as you answer. Claude's own menus, such as /model, and its feedback survey do not tell MAO they are open, so avoid leaving one open while tasks are arriving.
Fixes
- A failed turn no longer leaves an agent stranded. After an API error, a Claude Code agent could stop receiving tasks until you pressed Send.
Other agents work as before: Codex asks for Send on the first task of each session, and Gemini, OpenCode, Antigravity, custom commands and remote terminals ask every time.