Stop waiting to be asked: Snowflake's CoCo and CoWork are becoming always-on colleagues

On November 3, 2026, Snowflake announced a batch of agent capabilities at Expedition 2026, and the one most worth thinking about is called Snowflake Bots: give each of its two assistants, CoCo and CoWork, an always-on bot, hand it a goal, and it finishes the job in the background, coming to you only when a human decision is needed. It sounds like another "works automatically" marketing line, but Snowflake aimed this release at two questions that usually get skipped: on what basis does an agent know what counts as "right," and where do its permissions come from?

Start with the backdrop. Snowflake's official blog cites two numbers: Deloitte's 2026 Tech Trends report says only 11% of organizations have agents in production, despite 38% piloting them; Capgemini's research says trust in fully autonomous agents fell from 43% to 27% within a single year. Together, those numbers describe the real state of enterprise agents: demos look great, but against real data they fail — a model takes a plausible-looking table and confidently computes the wrong revenue number, or misses a policy constraint a junior analyst would have caught. Snowflake's diagnosis is that the problem is not whether the model is smart enough, but whether its context is authoritative enough: the agent does not know which metric definition is official, which data is current, or what permissions each user actually holds.

What Bots are: from "waiting for a question" to "owning a goal"

Snowflake Bots are positioned as goal-driven and always on. The usage pattern is: you give a bot a responsibility, such as "keep critical pipelines healthy" or "stay on top of customer churn signals," and it suggests which signals to monitor, plans the work, and carries the task through over days or even weeks, with data staying inside governed Snowflake and permissions carried over from yours. Its reporting rhythm is inverted: it does not wait for you to ask about progress; it comes to you only in two situations — when a decision needs human input, and when the job is done.

One calibration first: Bots are currently "private preview soon," not generally available. What is actually generally available (GA) in the same batch is CoCo Automations: a data engineer describes a recurring task in plain language — "check pipeline health before standup" or "rebuild the Monday report" — sets the cadence, and it runs on schedule, producing a transcript of each run, operating within the creator's permissions, and bounded by preset spending quotas. In other words, what you can use today is "work that runs on a schedule," while "a colleague that takes a goal and keeps at it for weeks" is still in line. Mixing those two tiers up is the easiest way to misread this announcement.

Why Snowflake can talk about permissions: the agent inherits your identity

The most important design sentence for Bots takes only half a line in the official blog: it works "under your existing permissions." Unpacked, that means the bot has no separate, larger identity of its own; it inherits exactly the access rights, role context, and permission boundaries of the user who assigned the task, and row-access policies and masking rules apply to it the same way. The same holds for CoCo's cloud sessions: every session runs under the creating user's Snowflake permissions, and administrators can tighten things further with restricted session scope.

This design targets the scenario enterprises fear most: an agent that starts running on its own at night and quietly reads a table it should not read or touches data it should not touch. Snowflake's answer is not another layer of after-the-fact auditing, but making overreach impossible from the start — what the agent can do equals what you yourself can do. Combined with a transcript for every run, its actions remain fully traceable afterwards. That may not excite indie developers, but for a company handing agents to hundreds of people, it is the difference between passing compliance review and not.

What changed for CoCo and CoWork each

CoCo, the coding agent for data teams, changed around "making progress while you are away." Besides Automations, Snowsight gained a CoCo mode: a full-screen workspace backed by a cloud container with a shell, file system, Git, and live preview, suited to long migrations, multi-step analyses, or pipeline builds — start a job, close the laptop, and come back to finished results. Cloud Agents are also generally available, letting teams run CoCo in a Snowflake-managed isolated environment for long-running tasks. CoCo is also moving into the places teams already work: a Slack bot (public preview), a GitHub integration (public preview, routing proposed changes through the team's existing pull-request process), mobile (iOS generally available, Android in public preview), and the Cortex Agent SDK (public preview) for embedding CoCo's coding capabilities in your own applications.

CoWork, the personal work agent for everyone else, changed around "tuning to each team." CoWork Profiles (public preview) let administrators configure which agents, skills, connectors, and instructions each team sees: a finance team's questions route automatically to a financial planning agent, a marketing team's to a campaign analytics agent, without users first picking from a menu of agents they do not recognize. Snowflake Dashboards (public preview) let anyone talk to a dashboard through CoWork — summarize it, ask what drove a change, filter to a segment — while the data behind it stays protected by row-access and masking policies. CoWork Automations, User Memory (GA), User Skills, MCP connectors, document generation (PDF, PPTX, DOCX, HTML), and HTML Reports (public preview) round out the picture, aiming to turn a conversation's output directly into shareable work products.

The layer underneath: teaching agents what "right" means

Bots and Automations solve the "doing" half; the other half of the announcement solves the "basis for thinking" half. Cortex Sense (public preview soon) auto-builds context from existing schemas, metric definitions, and business logic, saving weeks of manual setup. Analytical Search (public preview) can count, compare, and find trends across thousands of documents at once — for example, "how many contracts contain a price escalation clause?" The new Arctic Embed 3S embedding model (generally available soon) is claimed to improve retrieval quality over its predecessor by 3% in English, 6% in multilingual retrieval, and 23% for structured tables and columns, at the same compact size class. Cortex Training (public preview) lets teams fine-tune open-weight models such as Qwen, DeepSeek, Mistral, and Gemma on proprietary data, with Snowflake claiming costs up to 61% lower than competing serverless offerings — a vendor-measured figure, to be cited with care.

Two more ecosystem moves: the AI Data Package (public preview) bundles more than 60 curated external datasets with semantic views and a ready-to-use CoWork agent; Agentic Marketplace Discovery (public preview) lets users source data, apps, and AI products through conversation. Together with Skills and Plugins Sharing (GA), Snowflake's message is that one expert's proven workflow should become a reusable asset for the whole company, not something every team rebuilds.

My take: a release that puts the agent in a cage before letting it out

Set against recent agent announcements from OpenAI and Anthropic, the contrast is clear: others compete on how smart the model is and how high its task completion rate runs; Snowflake competes on whether an agent can be trusted to go out and act inside someone else's data sovereignty and permission system. Its bet is that the enterprise agent bottleneck is not the model, but context and governance — whoever builds "authoritative definitions, existing permissions, and complete audit trails" into the product first wins the production-environment ticket.

But mind the boundaries. First, Bots themselves are not open yet; what can be verified today is Automations and the surrounding integrations — do not let the "always-on colleague" framing convince you early. Second, figures like "up to 61% lower cost" or "23% better retrieval" are vendor measurements without independent verification. Third, Snowflake's whole logic assumes your data and permissions already live in Snowflake; if they do not, the elegant "permission inheritance" design has nothing to inherit from. For data teams, the pragmatic posture is: hand one or two genuinely recurring chores to the GA Automations, watch the transcripts and quotas for a month, and then decide whether to wait for the Bots private preview. For teams not on Snowflake, this announcement is better read as a buying reference — take its handling of "permission inheritance" and "auditability" and demand the same from your own agent vendor.

For teams without Snowflake: three things worth borrowing

Even if your data never touches Snowflake, three practices from this announcement transfer directly. First, the "permission inheritance" principle: before any agent or automation script acts, ask whose identity it runs under. Many small teams' scripts run on admin keys with full permissions, and only discover the blast radius when something goes wrong. Switching to least-privilege, per-task credentials is the cheapest security upgrade available. Second, the "transcript" habit: every automated job should leave a record that can be inspected afterwards — what ran, on what basis, with what result, at what cost — whether that is a log file, a daily digest, or an existing monitoring hook. Snowflake made transcripts standard for Automations; the habit ports to any automation. Third, tiered rollout discipline: your own product releases can copy Snowflake's three tiers — generally available, public preview, private preview. Many small teams ship in only two states, "live" and "not live," so a new feature launches at full volume and can only be rolled back at full volume. Opening a preview to a small group for two weeks and checking the data before full rollout needs no big-company infrastructure — a feature flag is enough.

One background figure for context: Snowflake's official release puts CoCo at more than 9,100 accounts and CoWork at more than 5,800, measured as weekly average usage over the last four weeks of the quarter ended July 31, 2026, naming Cisco and Docusign among organizations using both. Those are vendor figures without independent verification — cite the basis, not just the number.

Finally, a reading habit for any vendor's agent announcement: sort every new capability into "generally available, public preview, private preview soon" before setting expectations. In this announcement, Automations and the surrounding integrations are worth using now, Profiles, Dashboards, and the connectors are worth a small pilot, and Bots are worth watching until the private preview opens. Sorting by tier keeps "coming soon" from being mistaken for "already usable," and keeps genuinely usable GA features from sitting idle while you wait for something unreleased. Agents change fast; this way of reading announcements lasts.