Which Snowflake AI agents can work unattended now—and which remain previews?
Snowflake has packaged scheduled CoWork reports, CoCo data-engineering runs and forthcoming always-on Bots as one agent platform. The controls are specific, but the maturity labels are not uniform and no independent outcome study accompanies the launch.
The short version
- What happened
- CoWork can already rerun recurring reports under each user's permissions; Snowflake says CoCo automations and an asynchronous agent API are generally available, while its live CoCo documentation still labels hosted automation an open preview.
- Who it affects
- Data engineers may hand recurring diagnosis to cloud agents, while taking on more responsibility for permissions, durable instructions, trace review and production approval.
- What remains unproven
- All four sources are Snowflake-owned; no independent evaluator measured the new features.
Keep in mind
All four sources are Snowflake-owned; no independent evaluator measured the new features.
The public launch and live documentation conflict on whether CoCo Automations are generally available or in open preview.
Headline Bots are described as private preview soon, so their reliability and availability cannot be assessed from customer use.
The 5,800 and 9,100 adoption figures count accounts during an earlier four-week window, not individual workers, completed tasks or successful outcomes.
Customer quotations do not disclose matched baselines, sample sizes, error rates or review costs.
Availability differs by feature, cloud, region and regulated environment; a global press release does not establish access for a specific customer.
Multi-source analysis · 4 linked sources. Primary analysed source date: 3 November 2026. Earliest linked supporting source: 11 September 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 11 min The evidence, methods, implications and limitations
The direct answer: scheduled work is real, but the headline Bots are not generally available
Snowflake's 3 November launch combines several levels of automation under two renamed products. CoWork, formerly Snowflake Intelligence, can schedule recurring analysis and email a result; that function became generally available on 11 September. CoCo, formerly Cortex Code, can run data and software tasks from managed environments, record task history and continue after a person's computer is closed. The launch footnotes mark CoCo Automations and a new asynchronous agent API as generally available. Those are usable forms of unattended work, but they remain bounded by a saved prompt, a schedule and the user's permissions.
The more ambitious CoCo and CoWork Bots are different. Snowflake describes them as always-on agents that accept a high-level goal, plan and complete background work, act inside business tools and return to a person only for a critical decision or final sign-off. The same announcement labels Bots 'private preview soon'. The central vision is therefore a forward product commitment, not a capability available to ordinary customers today. Buyers should separate generally available recurring reports, hosted automation with a disputed maturity label and forthcoming goal-driven Bots.[1][3][4]
Why the product-status mismatch matters
Snowflake's launch footnote says CoCo Automations are generally available. Its live documentation, checked on the same day, titles the feature 'CoCo automations in CLI and Snowsight (Preview)' and calls it an open preview. That page says the feature is available in commercial regions across AWS, Azure and Google Cloud, but not in government, FedRAMP, Department of Defense, Virtual Private Snowflake or China deployments. It also lists preview limits: schedules cannot run more often than hourly, history is retained for two months, event-triggered runs are unsupported and a running task cannot be joined or resumed interactively.
A maturity label affects procurement, support expectations and risk approval. General availability normally indicates a supported production service; preview status often carries changing interfaces, restricted regions or different service commitments. The sources may reflect a documentation lag or a launch-day transition, but the public record does not resolve that. A buyer should obtain written, account-specific status rather than treating a press-release footnote as proof that every advertised surface and region is production-ready. CoWork's recurring-report feature is clearer because Snowflake published a dated general-availability release note in September.[1][3][4]
The distinction that matters
Snowflake's four maturity layers
The launch groups available products and future agents together; the operational difference is whether a person can deploy the capability now and what it is allowed to do.
CoWork reports
Rerun a question on a fixed schedule and email the result
Generally available since 11 September
CoCo automation
Run a saved data or coding task in a managed sandbox
Launch says GA; live documentation still says open preview
Connected action
Use business tools under the user's permissions
Availability varies by integration and control
Always-on Bots
Pursue high-level goals and surface selected decisions
Private preview soon; outcomes untested publicly
What an unattended CoCo run can actually reach
The CoCo automation documentation is specific about execution. Each automation becomes an agent task owned by the user who created it. When a schedule fires, Snowflake starts a managed sandbox, makes selected tools available, creates a conversation thread and records task state, timing, query ID and errors. The run uses the user's default role and default secondary roles, not necessarily the role active when the automation was created. A prompt that worked interactively can therefore fail, see less data or—if default privileges are broad—reach more objects than an administrator expected.
Snowflake grants the account-level EXECUTE AGENT TASK privilege to the PUBLIC role by default, while allowing administrators to revoke it and grant it more narrowly. Connected stages, secrets, GitHub credentials and Model Context Protocol servers require their own permissions. The company says every run leaves a trace, and its documentation recommends inspecting the conversation because a technically successful task does not prove the intended external action happened. These are useful controls. They do not measure whether the agent chose the right tools, interpreted data correctly or avoided an authorised but harmful action.[2][3]
CoWork expands automation from reports into business tools
CoWork's established automation model is narrower: it reruns a question with fresh data on an hourly, daily, weekly or monthly schedule and emails a summary plus a link to the full report. Snowflake says each run follows the caller's rights, respects row-access and masking policies, and stores the resulting conversation for up to 60 days. Email delivery requires a verified address. A forwarded report link does not grant another person access; sharing creates a snapshot. That design makes recurring analysis easier to review than an agent silently changing records across several systems.
The November package broadens the intended surface. Snowflake says CoWork can act through Slack, Gmail, Jira and Salesforce, while Teams, Outlook and Exchange can connect through custom MCP servers. It can save repeatable Skills, preserve user Memory and create Artifacts such as charts, documents and dashboards. Profiles can route teams to approved agents and data without widening underlying permissions. Several surrounding capabilities remain previews or are coming soon, including mobile experiences, profile functions, memory and auto-generated business context. Availability must be checked feature by feature rather than inferred from the CoWork name.[1][4]
The adoption figures are accounts, not proof of productivity
Snowflake reports that CoWork was used by more than 5,800 accounts and CoCo by more than 9,100. Its footnote says these were weekly account counts averaged across the final four weeks of the quarter ended 31 July 2026, covering capacity and on-demand accounts using the respective features. The denominator is accounts, not employees, active seats, completed tasks or paying organisations newly adopting the launch-day features. A single account may contain many users, and the figure does not reveal how frequently they used the tools or whether use continued after a trial.
Named customer statements describe faster pipeline conversion, easier analysis and reduced time from questions to answers, but Snowflake publishes no matched baseline, task sample, time distribution, error denominator or independent audit. The release does not report how many agent actions required correction, how often schedules failed, whether permissions blocked useful work, or whether labour time saved exceeded review and remediation time. The evidence supports a narrow conclusion: the products have material exposure across existing Snowflake accounts. It does not prove that they increase productivity for every knowledge worker or safely replace data-engineering work.[1]
Cost and control move with the worker's identity
Unattended work consumes tokens and compute after the person who created it has logged off. CoCo's documentation says preview automations incur standard Snowflake task billing plus token consumption. CoWork automations use the creator's warehouse and compute. Snowflake adds budgets, quotas, usage tagging and model routing intended to direct simpler work to less expensive models. The launch also advertises serverless model training and reports a best-case cost result up to 61% below selected alternatives, calculated from company-measured throughput on H200 GPUs and competitors' published October prices. That is a vendor benchmark, not a total-cost study across customer workloads.
Identity-linked execution creates both accountability and operational risk. A departing employee, changed default role or revoked dataset can alter later runs. A schedule may keep producing plausible reports from incomplete inputs if a permission changes silently. Administrators need ownership transfer, credential rotation, cost alerts, task inventories and a rule for disabling automations when roles change. Workers need to know whether an emailed report was generated from current data, which agent and model produced it, what actions it took and who remains responsible for checking the result.[1][3][4]
What changes for data engineers and knowledge workers
For data engineers, the practical change is that a coding assistant can move from an interactive laptop session into a managed cloud task that checks freshness, investigates pipeline failures or prepares a repair while the engineer is away. Snowflake says engineers retain production approval and can constrain a session's scope. If that works as intended, people may spend less time on recurring diagnosis and more on architecture, quality and exception handling. The work does not disappear: it shifts toward writing durable instructions, granting least privilege, reviewing traces and deciding when an automated fix is safe enough to merge or deploy.
For analysts and other knowledge workers, CoWork can turn a recurring question into a scheduled report and, over time, take actions through familiar tools. The human benefit depends on whether the report is accurate, timely and easier to verify than to recreate. It also depends on whether people can challenge a system that inherits institutional definitions and access policies. Memory and shared Skills may preserve useful context, but they can also preserve a mistaken assumption. Organisations should measure correction time, missed exceptions, permission failures and downstream decisions, not only the number of prompts or reports generated.[1][2][3]
What evidence would change the assessment
The first clarification should be a reconciled availability matrix showing exactly when CoCo Automations left preview, which regions and interfaces are covered, and what service commitments apply. Snowflake should separately date Bots, Memory, Profiles, mobile tools, spreadsheet integrations, model routing and Agent Studio. Buyers need the status for their own cloud, region and compliance boundary. Product pages that combine general availability, public preview, private preview and 'coming soon' under one narrative make it easy to mistake an announced direction for a deployable system.
The stronger evidence would be outcome data with denominators: completed unattended tasks, human interventions, false alarms, failed or incomplete side effects, access-control violations, security incidents, compute cost and time saved after review. Comparative evaluations should use matched workflows with and without the agent, disclose task difficulty and include serious failure cases rather than only average speed. Independent customer or auditor results would reduce reliance on vendor testimony. Until those data appear, the launch is consequential because it moves governed enterprise agents toward persistent work, not because Snowflake has proved that always-on AI coworkers reliably deliver better business outcomes.[1][2][3][4]
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 2 min 31 sec · Background explainer
What is artificial intelligence? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A short introduction to artificial intelligence from the Royal Society. Start here for basic concepts, then use current reporting to explore what has changed.
Evidence check: Published in 2018, before today’s generative AI tools. Useful for basic concepts, rather than an overview of current capabilities.
Published 2018-11-28. Source checked 14 October 2026. Check captions on the publisher’s video page.
Explore current AI technologyPhoto: ThisisEngineering / Unsplash · Licence
Continue the story
Read next
Continue exploring access control: compare the finding and its evidence limits.
AI agents retrieved the right memory—and still used a version they should not
Across 731 benchmark bases and six model providers, injected stale, missing, flipped or conflicting control state sharply reduced downstream scores. Oracle-backed repair recovered much of the loss, but it assumed the correct state was already known.
Read the next reportExplore the full source and evidence record
What this means for people
- Data engineers may hand recurring diagnosis to cloud agents, while taking on more responsibility for permissions, durable instructions, trace review and production approval.
- Knowledge workers can receive scheduled analyses and use agents across familiar tools, but still need visible provenance, current-data checks and a way to challenge incorrect institutional context.
- Administrators must manage user-owned schedules, default roles, secrets, billing and offboarding; a completed agent run is not proof that its intended business outcome occurred.
Global context
Enterprise software companies are moving assistants from chat into scheduled and tool-using work. Snowflake's distinctive claim is that the agent runs inside the data platform's existing roles, masking and audit systems. That may reduce the need for a parallel permission layer, but it also inherits every weakness in the organisation's roles, data definitions and offboarding. Comparable products use different identity, sandbox and approval models, so this launch cannot establish a general safety level for enterprise agents.
What the evidence does not yet show
- All four sources are Snowflake-owned; no independent evaluator measured the new features.
- The public launch and live documentation conflict on whether CoCo Automations are generally available or in open preview.
- Headline Bots are described as private preview soon, so their reliability and availability cannot be assessed from customer use.
- The 5,800 and 9,100 adoption figures count accounts during an earlier four-week window, not individual workers, completed tasks or successful outcomes.
- Customer quotations do not disclose matched baselines, sample sizes, error rates or review costs.
- Availability differs by feature, cloud, region and regulated environment; a global press release does not establish access for a specific customer.
What to watch next
- A dated Snowflake clarification reconciling the CoCo Automations GA and preview labels
- Actual availability dates and service terms for CoCo and CoWork Bots
- Independent task-level accuracy, intervention, failure, security and cost measurements
- Evidence that default-role execution and PUBLIC privilege grants are narrowed in real deployments
- Whether scheduled tasks expose complete traces and make incomplete side effects easy to detect
- Workforce outcomes including review burden, role changes and whether time savings persist after correction
Living evidence record
Impact record IAI-0VQSYLA
Evidence stage
Announced
Confidence
Supported
Reporting basis
Multi-source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
3 November 2026
Source trail
4 direct sources across 2 source types.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Evidence trail
Sources used for this report
Links checked 3 November 2026
Research question, topics and regions
Which parts of Snowflake's expanded CoCo and CoWork agent platform are available, preview-only or forthcoming, and what evidence exists about their effects
Topics: Enterprise agents · Unattended automation · Data engineering · Access control · Product maturity · Workplace AI
Regions: United States · Global
This report is labelled multi-source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.