Snowflake unveils CoCo and CoWork Bots: AI agents that keep working in the background after you log off

Snowflake announced on November 3, 2026, at its Snowflake Expedition 2026 event, a series of updates to its enterprise AI agents: Snowflake CoCo, described as the coding agent for data teams, and Snowflake CoWork, the personal work agent for knowledge workers. The company's core message is a shift from software that waits for an instruction to an autonomous layer that carries out work continuously, always within the platform's governance controls. This is a corporate press release, so claims about benefits come from the company itself and from customers it quotes.

The most striking new feature is CoCo and CoWork Bots, designed as always-on AI teammates. According to Snowflake, they act on the user's behalf and inherit the user's access permissions, role and context. The user sets a high-level goal, such as preparing briefings for every customer meeting or keeping data pipelines healthy. The bot plans and carries out the work in the background, within the user's governance boundaries, and only notifies the user when a decision or final approval is needed. The release says the Bots will soon be in private preview.

Another set of announcements covers automations. With CoCo Automations, now generally available, data teams can schedule recurring workflows with an audit log and task history, and can suspend or rerun an execution. CoWork offers something similar for business teams. For example, a sales manager can request an analysis of sales performance and schedule it to arrive in their inbox every week. In addition, the new Async Agent API, also generally available, lets CoWork handle long, multi-step tasks that continue in the background and resume if interrupted.

Snowflake is also bringing its agents into everyday tools. Snowflake for Slack combines CoCo and CoWork in a single bot experience. CoCo for Excel and Sheets brings Snowflake data and natural-language questions into spreadsheets, while the data remains protected in Snowflake. Snowflake Mobile is a unified app that lets business users query governed data and take action with approval, and lets developers follow coding sessions and manage builds from their phones. All three features are in public preview.

A significant part of the announcement focuses on making knowledge build up over time. Skills and Plugin Sharing, now generally available, lets users codify CoCo workflows and share them across the organization, with administrators controlling what counts as official. CoCo Profiles and the new CoWork Profiles, in public preview, provide context by role and team. In CoWork, administrators can set the agent's access and routing instructions that end users cannot remove. Memory remembers past decisions and context, and users can view, update or delete what is stored. One person's analyses, such as HTML reports or interactive dashboards, become discoverable by the rest of the team. Snowflake says it is bringing on the Mora team to strengthen this approach of applying AI to business intelligence.

On data and the marketplace, Snowflake introduced the Snowflake AI Data Package, in public preview, with more than 60 industry datasets, semantic views and a ready-to-use CoWork agent. It also introduced Agentic Discovery in Snowflake Marketplace, for finding and installing data, apps and AI through conversation. Providers can also generate a Cortex Agent and a semantic view for their listings so recipients can query them in natural language.

On cost, Cortex AI Gateway adds automatic model tiers with dynamic routing for CoCo, in private preview. The idea is to assign high-volume background jobs to lightweight models and reserve the most powerful reasoning for complex queries. Cortex Training, in public preview, lets customers adapt open-weight models to their own data with techniques such as reinforcement learning, on a serverless basis. Snowflake manages compute capacity and cuts costs through GPU multiplexing. The company says its benchmarks show up to 61% lower cost than other serverless training options. According to the release, the methodology is based on LoRA SFT workloads measured by throughput on an H200 GPU and on prices published by Fireworks and Tinker on October 22, 2026. The comparison was carried out by the vendor itself. Poetiq, a customer quoted in the release, says it uses Cortex Training for LoRA, SFT and full post-training with its GRPO implementation.

On adoption, Snowflake says more than 5,800 accounts use CoWork and more than 9,100 use CoCo, out of a base of more than 14,500 customers. It cites Cisco, Docusign, Klarent, Pacific Life Insurance Company, Shelter Mutual Insurance Company and United Rentals. Cisco talks about agents that spot upsell and renewal opportunities and about speeding up corporate data migration. Docusign mentions account research without needing SQL. United Rentals says tasks that used to take days now take minutes. These testimonials appear in the press release and have not been independently verified.

Overall, the announcement reflects Snowflake's push to position itself as a governed control plane for enterprise agents. Its argument is that an agent that doesn't know a company's definitions and rules can confidently give wrong answers. Several pieces, including the Bots, are not yet generally available, and the release includes the usual forward-looking statements about risks and uncertainties.