Where control is accumulating in the Agentic AI stack, and how to choose on purpose
For about fifteen years, the force that decided where enterprise value collected had a name. Dave McCrory called it data gravity in 2010, and the idea aged well.
Applications drift toward data because moving data is slow, costly, and risky. Whoever controlled the data layer controlled the decisions that stacked on top of it, from analytics to applications to budgets.
That logic still holds, but what reaches for your data has changed. The dashboards and pipelines that used to sit beside the warehouse are giving way to agents, and an agent does not stay next to the data. It runs on some platform, reasons over whatever it can reach, and moves results between systems on its own.
The platform you pick to run your agents is taking over the position the data layer used to hold. It is becoming the thing that owns the relationship with your data.
That is a bigger decision than it looks, and most teams are making it without noticing.
Once those walls go up, the position is expensive to win back, which is what makes this a decade-long bet and not a procurement round.
Agent gravity is the newer force
Tomasz Tunguz called this shift “agent gravity” in a recent essay. The argument runs parallel to the old one. Agents demand enormous compute, that compute is a large and growing business, and the platforms hosting agent workloads will fight to keep them. The more agents and data flowing through a platform, the heavier its pull.
Agents are turning into the main surface through which people and systems touch enterprise data.
An employee asks an agent instead of opening a dashboard. A customer interacts with an agent instead of a form. Other automated systems call an agent instead of hitting a database directly.
Once that becomes the default path, the platform running the agent sits closer to the value than the platform storing the data. Proximity to the work now accumulates more leverage than custody of the bytes.
Why running the agents is where the moat forms
Running an agent is expensive, and that expense is the point. Inference at scale, orchestration, memory, tool calls, retries, and the guardrails that stop an agent from doing something costly all burn compute, and compute is the business these platforms are in.
Tunguz has written separately about the harness, the orchestration and control layer that turns a raw model into something an enterprise can trust. That harness is where the hard engineering lives now. Whoever owns it owns the relationship with everything the agent reads, writes, and moves.
This is why the platform decision outlasts the model decision. Models will keep leapfrogging each other on every leaderboard.
The harness around them, the place your agents are configured, governed, and run, is sticky in a way individual models never were. I have seen an arrangement like this start as one convenient integration and end, two years later, with the bulk of a company’s analytical work running somewhere nobody picked on purpose.
The doors are already closing
Incumbents understand the dynamic, and they are not waiting for you to notice it. In April, Microsoft removed the compatibility mode that let Power BI query Databricks metric views through the standard connector, which broke the reports that relied on it (the release notes state it without ceremony).
At Build 2026, Microsoft positioned Fabric as the data platform for its Copilot and agent ecosystem, wired Fabric IQ into Microsoft 365 Copilot, and shipped Agent Skills that let agents build models and reports directly on governed Fabric data.
The behavior repeats across the field. Snowflake pushes Cortex, Google leans on BigQuery, and every one of them wants your agents reasoning over data inside its own walls.
The friction a vendor removes inside its own stack becomes friction everywhere else. That asymmetry is the gravity well, and it is built on purpose.
The question worth asking
This reframes what an evaluation should measure. Benchmark scores age in weeks, and a model that tops a chart today will sit mid-table by the next release. Tuning a ten-year decision around this quarter’s numbers misreads the timeline.
The operators I work with tend to ask a sharper question once they see the mechanics.
Which layer of the stack will own the relationship with our data over the next five to ten years? A company that stores its customer data in one system and runs its agents through another has already answered that question, whether it meant to or not. It handed the relationship to whoever controls the agent runtime, and it did so without holding a meeting about it.
Three checks separate a deliberate choice from an accidental one. First, can your agents read and write across platforms, or does every convenient path keep everything inside one vendor? Second, when an agent copies or moves data, who holds the audit trail and the off switch? Third, if you had to move your agent workloads to a different platform in three years, what would break, and what would it cost?
When the honest answer to the third question is that nobody has ever priced it, the platform has already priced it for you.
Make the bet on purpose
None of this argues for paralysis. Single-vendor stacks are convenient, and convenience earns its keep when a team is small and shipping fast. The narrower point is the one worth holding onto. The choice of where your agents run is compounding into control over your data, and that control is hard to win back once a vendor has built the gravity well around it.
Pick with open eyes, and price the exit before you need it. A platform decision you file under tactical has a habit of turning into the most strategic call you made all decade.
…
Nick Talwar is a CTO, ex-Microsoft, and a hands-on AI engineer who supports executives in navigating AI adoption. He shares insights on AI-first strategies to drive bottom-line impact.
→ Follow him on LinkedIn to catch his latest thoughts.
→ Subscribe to his free Substack for in-depth articles delivered straight to your inbox.
→ Watch the live session to see how leaders in highly regulated industries leverage AI to cut manual work and drive ROI.
Top comments (0)