One agent. You decide what it can do.
Talk through the work, let the agent use the tools you permit, and inspect what it remembers. It can carry a recipe across five tools while you watch. There is one agent, even when several providers do the work.
Announcement. Page 00 introduces the agent. There is one. Applications for a second position are not being accepted.
What the agent doesEight things, in the order you would meet them.
- Live
- Beta
- Insider preview
- Planned
- Status to confirm
Talk while the work continues.
Hear, think, speak and act run as four separate tracks. When you interrupt speech, the agent yields without discarding its thought. A held answer can resume. Tool work still passes through the normal permission checks.
Four tracks. One agent. The arithmetic is intentional.
Let it read without letting it act.
Each tool has separate Read and Control consent. Both start off. Read admits its live typing context; Control admits actions. You can allow or deny one artifact without changing the whole tool's grant. An explicit artifact deny wins.
Use the command catalog across your tools.
The catalog describes 36 modules and more than 750 commands. From home, the agent can list available action slices. Inside one module, it loads that module's slice.
Approve a destructive action before it runs.
An agent command marked destructive waits for your explicit yes. Decline or timeout returns a failure and records the outcome. Commands also have deadlines: 120 seconds for low-cost calls, 300 for medium, and 30 minutes for high-cost or streaming calls.
Read the sources behind a memory.
Open a memory summary and expand it back to its sources. The semantic store refuses memories without provenance. Forgetting one memory leaves a tombstone and erases its weights. Nightly consolidation turns episodes into semantic memories.
Plan locally when the connection goes away.
Local Gemma handles planning when the app is offline or you select local-first mode. That keeps a planning path available. External servers, hosted voice and tools that need a network still need a network.
Carry one recipe through five tools, with undo.
A typed recipe crosses C0:BRO, C0:CAP, C0:CHAT, C0:LINK and C0:TRK. Each step passes its output to the next. Real undo pairs compensate completed reversible steps on failure. Irreversible steps are surfaced, and outward actions still need approval.
Undo has a boundary. The recipe names it.
See where the agent acted.
The Mission Board puts promises, handoffs and chain steps on one time axis. Open a result to read what happened. Click Show to open the item or its module. Quiet stretches of at least 20 seconds fold into labelled breaks.
What the agent hands onAnd what it takes in.
- C0:DICThe agentThe live ear joins an existing dictation capture or owns a new one. It stops only the microphone it owns.Insider preview
- C0mpThe agentWorkspace facts, skills and citations feed bounded turns through the workspace binding.Insider preview
- The agentC0:CAPScene, viewport and selection reads support canvas variants for your approval.Insider preview
- The agentC0:CODEApproved canvas text, assets and spec cards become a build turn for an explicit website or app target.Insider preview
- The agentMCPCalls resolve to enabled connectors. An unknown or disabled connector returns a typed error.Insider preview
- The agentHistoryShape-only action traces record outcomes locally. Real failures also report to SRS.Insider preview
The same release research, twice.One job, with the handoffs left visible.
Separate apps
- Read the reference page in a browser.
- Put its screenshot on a whiteboard.
- Ask a chat assistant to summarize the material.
- Share the result in a messaging app.
- Record the work in a time tracker.
Five tools. You carry the results between them.
The one agent
- Review the five-tool recipe and grant the required Read and Control scopes.
- Run it and approve outward actions when asked.
- Inspect the results and the reversible steps on the Mission Board.
One recipe. The handoffs and undo boundaries remain visible.
Facts the lab is oddly proud of
- OffRead and Control both start off. An installed tool is not a permission grant.
- 7 daysAn intake question waits for three conversation turns, then asks at most once per seven days.
- TwoA learned recipe needs two confirmed successes and owner admission before replay. Success is not consent.
- SourcesA semantic memory without provenance is refused. Even a plausible one.
Not yetPlanned, and listed here so nobody has to guess.
- The Active Agent deskDesk, handover and attribution previews are building and disabled.
- Live own-key decision duet proofThe comparison and the scoreboard exist. A run with your own key has not been shown yet.
- Verified agent-to-agent deliverySigned envelopes and expiring consent exist. Real C0:LINK delivery and rendering need separate proof.
- Gateway and routing settings that take effectGatewayMode, ModelRoutingPrefs and ReducedMotionStage are parked. Their consumers are not wired.

