Oleksandr Kuzmenko, EditorPublished 7 min readAI-assisted · editor-reviewed
TL;DR
- Assign dedicated service identities to AI agents rather than reusing personal user OAuth tokens to maintain strict audit trails.
- Partition agent memory into shared team context versus isolated member-specific attributes to avoid cross-polluting contexts.
- Run background tool agents inside sandboxed cloud compute instances rather than granting unrestricted local access.
Key facts
- Group limit
- Up to 6 members
- Agentic harness
- Google Antigravity
- Core models
- Latest Gemini models
- Availability
- US personal Google accounts (18+)
Isolated Cloud Runtimes and Verified Agent Identities
Google Labs has opened access to CC, an experimental collaborative agent designed for groups of up to six people. Rather than running locally or impersonating a specific user through delegated session tokens, each CC instance runs on a dedicated cloud computer managed by Google's proprietary agentic harness, Antigravity. The agent operates under its own verified Google Account, creating a clear audit log and security boundary for actions taken across integrated services including Gmail, Google Chat, Docs, and Calendar.
Dual-Tier Memory and Selective Context Ingestion
To prevent context pollution and privacy leaks between participants, CC introduces partitioned memory architecture:
- Shared Group Memory: Stores group-wide operational context, including shared schedules, vendor preferences, and recurring task lists.
- Individual Attribute Storage: Maintains distinct user profiles (such as dietary restrictions and individual timezone constraints) without exposing raw personal messages to other group members.
- Explicit Ingestion Hooks: Data enters the agent's context through selective allowlists—such as designated sender filters for emails—or direct uploads of registration PDFs and schedule snapshots via Google Chat or Google Drive.
Production Blueprint for Multi-User Agents
By executing inside an isolated cloud environment, CC interacts with external tools (such as live travel time queries via Google Maps API and automated PDF form pre-filling) without placing credentials on client hardware. The agent is rolling out to personal Google accounts in the United States via a waitlist and upgrade invites.
When to use
- When designing agentic systems that coordinate schedules, documents, and notifications across multiple collaborating users.
- When researching patterns for giving autonomous agents distinct service identities with explicit permission boundaries.
When NOT to use
- When building single-tenant local coding agents where low latency and direct filesystem execution take precedence.
- When handling enterprise sensitive codebases requiring strict on-premise execution and zero cloud data sharing.
What to do today
- Separate agent credentials from personal OAuth tokens by generating dedicated bot service accounts in your agent architectures.
- Implement a two-tier memory store separating global team metadata from user-specific attributes in agent frameworks.
- Sign up on the Google Labs CC waitlist to evaluate multi-user agent interaction workflows.
Step names are taken from What to do today.
Step names are taken from What to do today. Separate agent credentials from personal OAuth tokens by generating dedicated bot service accounts in your agent architectures.. Implement a two-tier memory store separating global team metadata from user-specific attributes in agent frameworks.. Sign up on the Google Labs CC waitlist to evaluate multi-user agent interaction workflows.
- Separate agent credentials from personal OAuth tokens by generating dedicated bot service accounts in your agent architectures.
- Implement a two-tier memory store separating global team metadata from user-specific attributes in agent frameworks.
- Sign up on the Google Labs CC waitlist to evaluate multi-user agent interaction workflows.
What the community says
This actually looks a lot like something I was looking for. Sadly, it’s from one of the last companies I’d be willing to allow access to more of my data.
Interesting choice with the name. Claude Code is sometimes abbreviated to CC.