Rabbit starts closed beta of RabbitOS 3, a top-level agent that coordinates other agents, skills, and devices instead of adding standalone features.
The AMW Read
RabbitOS 3 meaningfully advances the agent segment's UX baseline toward orchestration-over-skill-catalogs, but it is a known player's closed beta with no funding, entrant, or debate-resolving event disclosed.
Rabbit starts closed beta of RabbitOS 3, a top-level agent that coordinates other agents, skills, and devices instead of adding standalone features.
Rabbit has moved RabbitOS 3 into small-scale closed testing, according to a hands-on account published by PingWest. The system sits above other agents, skills, models, and connected devices: a user states an intent in a single chat interface, and RabbitOS 3 decides which sub-agent to invoke, which skill to install, and which device to operate. Device setup is reduced to pasting one line of code into a terminal to register a Windows, Mac, or Linux machine as a controllable node under one account, and skills are installed by pasting a GitHub link rather than following manual dependency and config steps. The product is described as natively compatible with OpenClaw, Hermes, and other GitHub-hosted skills, drops session-based chat tabs for one continuous stream with persistent shared memory, and generates interface elements — tables, download buttons, payment confirmations — dynamically per task.
The release targets a friction point that has kept agent adoption concentrated among developers: install steps, permission grants, and skill configuration most users won't complete unassisted. Framing RabbitOS 3 as a coordination layer over third-party skills and models, rather than a closed catalog, sets Rabbit's distribution bet against the session-per-task, skill-per-app pattern that defines most current agent products. Rabbit reports trained tasks running 50 to 300 times faster than screenshot-based computer-use approaches — a self-reported figure from its own benchmark, not independently verified.
For builders, one-click import of existing OpenClaw-style configurations lowers switching costs for developers already running agent workflows, though that wedge only holds if third-party skill authors keep targeting compatible formats. For investors, the signal to watch isn't a new capability claim but an interface bet: whether users want one account-bound orchestrator spanning every device — PC, phone, eventually car and smart home — over app-specific agents, a question the closed beta won't answer until usage scales.