- MODEL CLASS
- Vision-language-action agent
- plans, calls tools, and verifies its own output
- CONTEXT
- Long-context window
- holds full task histories in-frame with retrieval fallback
- MEMORY
- Persistent across sessions
- resumes paused tasks without a re-prompt
- TOOLS
- Open connector set
- browser, files, APIs, and community-built adapters
- VOICE
- In-house speech stack
- lower hallucination rate, faster time-to-first-token
- LATENCY
- Sub-second first token
- streaming plan visible before the agent acts
- AUTONOMY LOOP
- Plan → act → verify → log
- every step written to an inspectable record
- EVAL
- Open, reproducible suite
- community submits tasks; leaderboard posts automatically
- RUNTIME TARGET
- 1M+ autonomous actions
- and counting, in public
BRAINPlanning and reasoning core
A vision-language-action model turns a plain-English goal into an ordered plan, then narrates each step before acting so watchers can stop it mid-run.
HANDSTool-use and execution layer
An open connector set lets the agent touch the real world — browser, files, and APIs — with retries and self-verification wired into every call.
PACKMemory and identity
Persistent memory carries context across sessions, and the chrome cat-head identity frame keeps every run recognizably Artificial Inu.