How Maia Works With Snowflake CoCo and CoWork

The Complete Agentic Data Workflow

Snowflake handles agentic consumption. Maia handles agentic supply. Here is how the two compose, stage by stage and handoff by handoff.

At Snowflake Summit 2026, Snowflake shipped the consumption side of the agentic enterprise. CoWork gives every business user a personal work agent, CoCo gives every developer a coding agent, and Horizon Catalog governs what both can touch. What Snowflake did not ship, deliberately, is the autonomous team that builds and maintains the data products all those agents consume.

That is what Maia does. Maia is a team of specialised data engineering agents covering design, ingestion, transformation, testing, deployment, monitoring, and governance. The agents operate against your team's standards in the Context Engine and are supervised through Mission Control. Maia runs on Snowflake compute, deploys from Snowflake Marketplace, and connects to Snowflake's agents through the protocol layer announced at Summit.

This piece is a practical guide, covering how these platforms work together, how a request moves between the two platforms, and what the architecture looks like underneath.

The Division of Labour

The cleanest way to understand the partnership is by workflow stage.

Snowflake owns the consumption layer.CoWork answers business questions over governed, live data. CoCo helps engineers write SQL, dbt, and Python inside the Snowflake environment. Cortex Sense auto-gathers context to make agent answers accurate. Horizon Catalog enforces what every human and every agent can see and do.

Maia owns the supply layer.When the data product a CoWork agent needs does not exist yet, Maia's agents build it. The design agent produces the model. Mission Control orchestrates the work as a sequence of tasks: connecting the source from Maia's library of 130+ connectors, applying your architecture standards during transformation, running SLA and quality checks, and deploying to production with scheduling, monitoring, and lineage attached. Human-in-the-loop checkpoints sit between each stage, so an engineer signs off before the pipeline moves forward. After go-live, Maia keeps operating the pipeline: detecting schema drift, monitoring data quality, and remediating issues on its own.

Two components make the supply side governable.Context Engine holds your team's standards: naming conventions, architecture patterns, data contracts, business definitions. It's encoded once and read at every stage of the pipeline, so the same conventions govern ingestion, transformation, testing, and deployment alike. That consistency is what makes it possible to hand work between stages, and between Maia and Snowflake's agents, without re-explaining the rules each time.

The relationship with CoCo is complementary, and the boundary is practical. CoCo is the right tool when an engineer wants to write or explore code conversationally inside Snowflake. Maia is the right tool when the team needs pipelines designed, built, tested, deployed, and operated in parallel, which is work nobody wants to do one prompt at a time. Many joint customers will run both.

How a Request Actually Moves

Here is the workflow with both platforms in place, walked through three scenarios joint customers hit in week one.

Scenario 1: The Business Asks a Question the Data Can't Answer Yet

A regional sales lead asks their CoWork agent for margin by product line, blended with a supplier dataset that has never been ingested. CoWork can answer the first half. The second half is a supply request. That requirement routes to Maia, via the data team's queue today and increasingly via direct agent-to-agent handoff as Snowflake's MCP and ACP support rolls out. Maia translates the business requirement into a data product spec and surfaces the plan in Mission Control, where an engineer reviews and approves the approach before any work starts. From there, Maia connects the supplier source, builds the transformation against the standards in the Context Engine, and validates it with automated tests. The finished pipeline comes back to Mission Control for a second review, this time of the actual product, not just the plan, and the engineer approves it for release. The new product is published into Snowflake with lineage Horizon can read, and the sales lead's next CoWork question lands on it. The request never queued behind a sprint.

Scenario 2: A Source Schema Changes Overnight

A source system adds a column and renames a field at 2am. Without automation, the pipeline fails, a dashboard breaks, and an engineer spends the morning investigating. With Maia operating the supply layer, drift is detected automatically, the affected pipeline is rebuilt against the conventions in the Context Engine, and SLA checks re-run before the artifact is republished. By morning, CoWork agents across the business are pulling from current data, and Mission Control shows the Head of Data what changed, what was rebuilt, and what passed. Nobody got paged.

Scenario 3: A Legacy Estate Is Blocking the Rollout

Most enterprises planning a CoWork deployment hit the same wall. The data they need is produced by decades-old Informatica or SSIS pipelines running outside Snowflake. Snowflake's AI-powered Migrations move the warehouse; they do not move those pipelines. Maia does. It ingests the legacy job XML, reverse-engineers the business logic, generates equivalent Snowflake-native pipelines, validates them, and ships with lineage from day one. Maia converted 100 Informatica pipelines in 30 minutes in a live public webinar earlier this year. Every migrated workload becomes governed data inside Snowflake, available to CoWork and CoCo on arrival.

There is a fourth pattern worth naming because it gates everything else: the semantic model. CoWork is only as accurate as the layer mapping business terms to physical tables, and building that layer is normally a quarter of analytics engineering work. Maia's Context Engine, paired with a purpose-built semantic modeling skill, generates a Snowflake-compatible semantic model directly and keeps it aligned as sources evolve. CoWork launches on a complete semantic layer instead of a partial one.

The Architecture Underneath

Three mechanics make the composition work rather than just sound good.

Push-down execution keeps everything inside the warehouse.Maia's pipelines execute on Snowflake compute. There is no external processing engine, no data leaving the governed perimeter, and no lineage break. Every pipeline Maia builds and operates runs inside Snowflake, visible to Horizon.

Protocol-level handoffs connect the two agent teams.MCP, ACP, and Cloud Agents are the protocol layer Snowflake announced at Summit, and they are the integration points that let an agent inside Snowflake and an agent outside Snowflake compose without either owning the other. A CoCo session can hand a supply request to Maia. A Maia agent can hand a finished data product back to CoWork.

Marketplace procurement removes the buying friction.Maia deploys directly from Snowflake Marketplace against existing committed spend, with no new procurement motion. Landkreditt Bank deployed this way and cut new data source onboarding from months to days.

The partnership behind the architecture is long-standing: Snowflake Ventures invested in Matillion in 2022. Sridhar Ramaswamy, Snowflake's CEO, put it this way: "Maia offers a glimpse into the future of data engineering. It's intuitive, powerful, and feels like a real accelerant for how teams build with data."

What Joint Customers See‍

Balfour Beatty watched Maia parse legacy Informatica XML in 6 minutes that had taken a senior engineer a week, turning a stalled multi-year migration into a 6-month delivery window across 1,300 pipelines.

At Precision Medicine Group, the combination carries clinical weight. "Our ability to generate trustworthy data is meaningful because the results, leveraging Maia and Snowflake, are being used to drive drug approvals offered to patients," says Roberto Lara, VP Digital Transformation & Analytics.

The pattern across all of them is the same. Snowflake's agents make the business faster at asking. Maia's agents make the data team faster at supplying. Put together, a business requirement travels from a question in CoWork to a governed data product in production without queueing behind a human at every step.