The Agentic Enterprise Still Runs on Human Supply

Snowflake just made its builders faster. The next step is making the workflow autonomous.

A POV on the Snowflake Summit 2026 product keynote, and where I think Snowflake customers go from here.

A day after Sridhar Ramaswamy made the strategic case for the agentic enterprise, Christian Kleinermann and Benoit Dageville walked through the product layer that operationalizes it. The keynote was wide-ranging; almost every part of the stack received a new capability or public preview, and the engineering work is impressive. What I want to spend my time on is what the pattern of investment tells us about where the workflow goes next.

The short answer: Snowflake just made consumption agentic. The next constraint is supply.

What Snowflake shipped, and what it means

The announcements cluster into four areas.

Personal agents. CoWork (formerly Snowflake Intelligence) puts a personal work agent in front of every business user, with multi-agent orchestration, scheduled tasks, and certified shared artifacts over live data. CoCo (formerly Cortex Code) does the same for developers, GA today across CLI, VS Code, Cursor, Excel, and Desktop, with a roadmap running through Snowsight, Airflow, dbt, MCP and ACP support, and Cloud Agents.

Agent infrastructure. Cortex Sense lifts agent accuracy on Snowflake's own eval from 24% to 83% by auto-gathering context. Horizon Catalog adds intent-driven governance, agent identity as a first-class concept distinct from user identity, and multi-party approvals. These are the rails the personal agents run on.

Ingestion. OpenFlow gains programmatic APIs and new connectors including Oracle GA, Mongo, and Shopify; zero-copy partnerships expand to Workday, IBM Watson Next Data, and SAP now GA; and the Natoma acquisition wires CoWork and CoCo into over a hundred business systems.

Modernization. AI-powered Migrations and Teradata virtualization make it materially easier to move off legacy warehouses and onto Snowflake.

Stack those together and the thesis is clear: Snowflake is making every person who works inside the platform — developer, analyst, or business user — meaningfully faster. That shows up immediately in productivity for teams whose data is already well-modeled inside Snowflake.

The question I keep coming back to is what happens at the edge of that productivity gain.

Amdahl's Law, applied to the data workflow

Amdahl's Law is a piece of workflow physics. It says that when you speed up just one stage of a workflow, your total speedup is capped — and the cap is set by the stages you left alone. Halve the time of the stage you optimized and you barely move the total; the untouched stage now dominates. The faster the optimized stage gets, the more the slow stage becomes the thing everyone waits on.

Apply that here. Snowflake just made the consumption stage fast and agentic. CoWork makes business users faster at asking questions. CoCo makes developers faster at building. Cortex Sense makes agents more accurate. These are real gains.

But the workflow does not start with consumption. It starts with supply: getting the right data into Snowflake, in the right form, with the right governance, against the right business requirement. That stage is still largely human-led, even when the human has better tools. A data engineer still configures pipelines, maps fields against medallion conventions, sets policies, monitors schema drift, and handles change.

So the workflow physics tells us where the next constraint lands. The faster Snowflake makes consumption, the more the supply stage becomes the bottleneck on how fast a data team can ship. This is not a criticism of what Snowflake announced — it is the natural arc of workflow optimization: solve one stage, and the next one comes into focus.

The largest near-term opportunity for Snowflake customers, I think, is bringing the same agentic shift to supply.

How Maia extends what Snowflake shipped

Maia is complementary to what Snowflake shipped, not competitive with it — it operates on the opposite side of the workflow. Maia delivers a team of specialized autonomous data engineering agents — covering design, ingestion, transformation, testing, deployment, monitoring, and governance — operating against a team's standards encoded in the Context Engine and supervised through Mission Control. What makes the composition with Snowflake work is the protocol layer Snowflake is building. MCP, ACP, and Cloud Agents are exactly the integration points an autonomous supply-side team needs. A CoCo session can hand a supply request to Maia. A Maia agent can hand a freshly-built data product back to CoWork. Snowflake is making the builder tooling inside the warehouse more capable; Maia is making the supply work outside the warehouse autonomous. Together the customer gets an end-to-end agentic workflow rather than agentic consumption sitting on top of human-led supply.

Here is how I think that plays out in practice.

CoWork drives demand. Maia keeps supply current. Picture a federated business where a central data engineering team supports users across regions and functions. The pattern I keep seeing is the central team becoming the bottleneck as demand for data outruns what they can serve. Roll CoWork into that picture and every federated team gets a personal agent pulling fresh insights — but only as fast as the underlying data stays current. When a Maia agent detects an overnight schema change, it rebuilds the affected pipeline against the conventions encoded in the Context Engine, the testing agent re-runs SLA checks before the artifact republishes, and Mission Control gives the Head of Data a single view of lineage, freshness, and policy compliance. The federated team that used to wait on a central queue now self-serves in CoWork, because the supply got automated underneath them.

Maia builds the semantic model. CoWork uses it from day one. CoWork is only as accurate as the semantic layer underneath it. Without one that maps business terms to physical tables, the agent has rows and columns but not concepts and metrics, and answers degrade fast. The Cortex Sense jump from 24% to 83% is the proof point: context is the whole game. The catch is that someone has to build the semantic model first — and today that someone is a data engineer or analytics engineer working with the business to encode the entities, write the joins, define the metric logic, document the business glossary, and keep all of it in sync as upstream tables evolve and new questions show up. The teams I talk to either ship a partial model and accept that CoWork answers half the questions reliably, or block the rollout for a quarter while a proper model gets stood up. With Maia, the design agent produces the semantic model CoWork needs, the transformation agent keeps the underlying products aligned, and the governance agent applies policy. Standing CoWork up on a clean semantic layer goes from a quarter to a day.

Maia migrates the pipelines. AIM migrates the destination. Snowflake's AI-powered Migrations make it faster to move data off legacy warehouses. They do not move the pipelines that produced that data: the ingestion jobs, transformation logic, orchestration, and tests that fed the legacy estate still belong to the customer's engineering team. For most enterprises I talk to, that legacy ETL estate is the actual bottleneck on going agentic. An Informatica or Oracle Data Integrator estate has decades of accumulated business logic buried in undocumented jobs; a senior engineer can take a week to reverse-engineer a single pipeline, turning migration into a multi-year program that blocks the CoWork and CoCo rollout entirely. With Maia's Migration Agent, that calculus changes. Maia ingests the legacy job XML, reverse-engineers the business logic against standards in the Context Engine, generates equivalent Snowflake-native pipelines, runs validation, and ships with Horizon-friendly lineage from day one. Maia demonstrated 100 Informatica pipelines converted in 30 minutes earlier this year. Multi-year programs compress to multi-quarter.

Where I think this goes next

I want to come back to the architecture point, because I think it is the most consequential part of what was announced. MCP. ACP. Cloud Agents. These are the protocol layers that let an agentic AI system inside Snowflake and an agent outside Snowflake compose without either having to own the other. They say the future of the agentic data workflow is composable, not monolithic. That is a significant architectural bet, and I think it is the right one.

For data leaders planning the next twelve months, here is my read. The investments Snowflake announced will pay off immediately for the data you already have well-modeled inside the platform. The next step — moving from agentic consumption to agentic end-to-end — is composing what Snowflake shipped with an autonomous supply-side team. That is where the biggest near-term gain is. And the connection points to get there are already in place.