Snowflake 为 CoCo 和 CoWork 推出常驻后台智能体

Snowflake announced on November 3, 2026, new always-on bots for CoCo, its coding agent for data teams, and CoWork, its personal work agent for knowledge workers. The bots are designed to plan and complete work in the background rather than wait for individual prompts.

The CoCo and CoWork Bots are coming soon to private preview. A user can set a high-level goal, such as preparing briefings for every customer meeting or keeping data pipelines healthy. The bot inherits that user’s access rights, role context and permissions, works within those boundaries and surfaces the work only when a critical decision or final approval is required. That design makes the approval boundary central to the product: Snowflake is extending agents from answering requests to pursuing continuing assignments, while leaving existing access controls and human sign-off in place.

Scheduled Work and Everyday Tools

CoCo Automations is generally available, allowing data teams to schedule repeated workflows through CoCo and CoWork. Each run retains an audit trail and task history, and teams can suspend or rerun a job. For business users, CoWork can repeat an analysis on a set schedule; Snowflake’s example is a sales leader receiving a weekly report on sales performance, emerging trends and areas needing attention. The generally available Async Agent API supports longer, multi-step CoWork tasks that continue in the background and resume after an interruption.

Snowflake for Slack, CoCo for Excel and Sheets, and Snowflake Mobile are in public preview. Slack places CoCo and CoWork in one bot, while the spreadsheet integrations let users ask plain-language questions while the data remains protected in Snowflake. Snowflake Mobile combines both agents, giving business users a way to approve actions and letting developers continue coding sessions, manage builds and steer work from their phones.

Governance, Memory and Reusable Work

Other controls address what happens when agent use spreads beyond an individual. Skills and Plugin Sharing, now generally available, lets users turn CoCo workflows into reusable assets and share them across an organization; administrators can decide which workflows are designated official. Profiles in CoCo and the new Profiles in CoWork are in public preview. CoCo Profiles offer suggested defaults that users can customize, while CoWork Profiles let administrators set agent access and routing instructions that end users cannot remove.

Memory in both agents retains previous context and decisions, with users able to view, update or delete stored memories. CoWork can also turn analysis into shareable HTML reports or interactive dashboards, and teammates can open a side-by-side conversation with CoCo to examine the results. Together, those features connect personal history and team reuse: context persists, but people and administrators retain defined controls over what is remembered and which workflows become standard.

Adoption and Customer Deployments

Snowflake said CoWork is used by more than 5,800 accounts and CoCo by more than 9,100. Those figures are averages over the last four weeks of the quarter ended July 31, 2026, counting capacity and on-demand accounts that used the respective features weekly. Cisco, Docusign and United Rentals are among the organizations using both products.

Cisco said its Customer Experience teams use specialized agents built in Snowflake to identify upsell and renewal opportunities, while CoCo helps engineers convert legacy code and build pipelines for migrating core corporate data.

Docusign said CoWork powers conversational data agents that let go-to-market teams perform account research and obtain insights without SQL expertise.

The announcement ties autonomous work to a chain of controls: inherited permissions for bots, administrator rules for shared workflows, approval checkpoints for consequential decisions and audit trails for scheduled jobs. That makes governance part of the operating mechanism rather than a separate policy layer. With the bots still awaiting private preview, the practical test is whether customers can extend the available automation and sharing features into continuous work without losing visibility into what the agents did.