For the past few years, the hard work in AI has gone into models. Organizations spent that time learning, experimenting, and building the best models they could. That work paid off, and it cleared the way for what’s happening now, everywhere, at breakneck speed: agents.
Companies have found real uses for agents across the organization, from customer support to insurance claims and autonomous vehicles. But while the attention stays on building the smartest models and agents, there is a quieter, yet gigantic problem underneath. These autonomous, goal-driven agents read, write, and transform enterprise data on their own, at machine speed, with no human reviewing each action. That changes what the data layer has to handle, and it will make or break the success of agentic AI.
The pressure behind that is real. According to Dun & Bradstreet’s recent AI Momentum Survey, 97% of organizations report active AI initiatives, but only 5% say their data is adequately ready to support them. That gap is the backdrop for everything in AI right now. Agents don’t close it. They widen it.
This week, we announced lakeFS for Agentic AI. It brings governed, reproducible data access to autonomous and headless agent workloads. It runs on the same scalable data version control architecture our customers already run in production, and gives tools, users, and agents one control plane for AI-ready data.
Why agents make the data problem worse
The trouble starts when agents access data, including images, video, metadata, and structured tables, without proper guardrails and at scale. The manual governance and operational controls most teams built for human-driven workflows were never designed for that.
The risk is simple to state. Any agent that reads or writes to production data without isolation or a reproducible trail is a liability, no matter how good the underlying model is. When an agent makes a mistake, you need to know what it touched, replay exactly what happened, and roll back without taking production down with it. Most stacks can’t do that today.
How lakeFS solves it at the data layer
lakeFS gives every agent its own isolated data sandbox: a zero-copy branch of the data it needs. The agent works there, lakeFS validates and merges its changes under policy, and every action lands in a unified audit trail. Four things make that work.
Isolation
Every agent gets a zero-copy branch covering structured tables, unstructured files, and metadata together as one. Agent mistakes are isolated automatically and never corrupt production data. Recovery that used to take hours takes seconds.
Reproducibility
Every agent run is tied to an exact, immutable version of the data. You can re-create, debug, audit, or extend any past action using the same inputs.
Governance and compliance by design
Production data is gated by policy. Merges happen only after pre-merge validations pass, and every change can carry agent identity, run ID, and execution context. You get one audit trail instead of evidence scattered across orchestrators, model providers, and cloud logs.
Agent-native infrastructure
Agents read and write through standard file operations, with branch-scoped credentials that confine each agent to its own workspace. That keeps each agent’s working set narrow and avoids context bloat. No custom MCP server, SDK, or specialized integration required.
The companies that win will solve this at the data layer
Agentic AI is moving from pilot to production, and the teams that treat agents like the production workloads they are, with the isolation, reproducibility, and audit trail that implies, are the ones who will end up trusting agents with real data. The rest will keep agents stuck in demos because nobody can answer the basic question of what the agent touched and whether it’s safe to let it run.
lakeFS for Agentic AI is available today to every lakeFS Enterprise customer, built on the same architecture you’re already running. If you have agents in flight, this is ready when you are. Read the full lakeFS for Agentic AI announcement.
Come dig into this with us
If this is the problem on your desk right now, join Oz Katz, our CTO and Co-Founder, for a live webinar on June 24, 2026 at 11 a.m. ET:
“When Agents Touch Production Data: Trust, Isolation and Reproducibility for Agentic AI.“
