Snowflake’s November 3 Expedition 2026 announcement puts CoCo, its coding agent for data teams, and CoWork, its work agent for knowledge workers, on a path from conversational assistance toward recurring work. The launch combines background-running Bots, scheduled CoCo automations, and tools intended to give agents business context and governance. But the features have different availability statuses, and Snowflake’s product announcements should not be mistaken for independent evidence of how well they perform.
11
19
What changes with CoCo and CoWork
Snowflake says the updated agents can take on recurring tasks rather than wait for every step to be prompted. Its new CoCo and CoWork Bots are described as working in the background and checking in when a human decision or sign-off is needed.
11
13 Coverage also reports that Bots can pursue assigned goals and inherit employees’ access rights, keeping access tied to existing permissions.
18
23
Examples reported around the launch include preparing a customer-meeting briefing and handling ongoing work. For data teams, Snowflake separately describes CoCo automations for tasks such as checking pipeline health or investigating failed jobs.
21
24 These are product use cases, not evidence that every task runs without human review.
CoCo Automations and spreadsheet access
CoCo Automations let data teams describe recurring tasks in plain language and have them run proactively. Snowflake lists the capability as generally available in Snowsight and the command-line interface, with public preview in CoCo Desktop.
24 The available announcement details support scheduled automation, but do not establish the specific per-run audit-trail claim.
Snowflake also listed CoCo for Excel and CoCo for Google Sheets as preview features in its release notes.
2 The supplied release information does not confirm the requested preview status for Snowflake for Slack or a unified Snowflake Mobile app, so those should not be treated as established launch details here.
Business context and controls
The agents are part of a broader effort to make enterprise AI act with shared business context and centrally managed controls. Snowflake’s Business Ontology is intended to provide agents with shared, preapproved business definitions; coverage describes it as building on Snowflake’s semantic modeling capabilities.
14
Cortex AI Gateway inference, observability, and cost management became generally available on November 3. Snowflake’s release notes describe the gateway as a governed endpoint for organizational AI clients that records request traces and attributes spend to the users who made the requests.
8 That provides a control layer around AI access; it does not, by itself, establish that an agent’s decisions are correct.
Adoption and a separate model-training announcement
Snowflake’s fiscal Q2 results reported more than 9,100 CoCo accounts and 5,800 CoWork accounts.
38 These figures offer adoption context, but the cited results do not establish the four-week averaging definition or mean that each account represents one individual user.
Separately, Snowflake announced Cortex Training in public preview for post-training open-weight models, including Qwen, DeepSeek, Mistral, and Gemma, using managed GPUs.
17 Snowflake also claims its LoRA benchmarks showed up to 61% lower cost than competing serverless training APIs, including Tinker and others.
37 That is a vendor-reported, best-case benchmark claim—not a general savings guarantee or an independently verified comparison.
What to watch
The core shift is clear: Snowflake wants its agents to handle repeatable work while using company-specific definitions and existing access controls. For buyers, the practical questions are which features are available in their chosen interface, what actions still require approval, and how each automated run can be monitored. The launch materials describe the intended capabilities, but do not independently demonstrate their reliability or customer outcomes.