The Intelligent Data Platform for the AI Era

Some of you may have noticed a different, and perhaps familiar, name at the top of this post.

Following this week’s leadership change, the board asked me to step in as interim CEO. I led MongoDB for more than 11 years and have served on the board ever since, so I've stayed close to our people, our customers, and the business. I’m happy to be back, and excited to share where MongoDB is headed next.

AI is undergoing an important transition. Models are getting better at reasoning, but the bigger change is what we’re asking them to do. AI is moving from generating answers to taking action. Software no longer just helps people do their work. Increasingly, agents do parts of the work themselves. They resolve customer issues, write and deploy software, execute business processes, and operate continuously with progressively less human intervention.

That shift has profound implications for the infrastructure underneath applications. An agent doesn’t simply take a prompt and return an answer. It observes what is happening, retrieves context, reasons about what to do, takes action, sees the result, and often starts the cycle again. As agents become more capable, these loops become longer, more autonomous, and more consequential.

Model quality matters enormously. But so does the quality, freshness, and accessibility of the data the model operates on. That is where MongoDB plays an important role.

Today, we’re announcing the largest expansion of our platform yet:

• MongoDB 9.0 is the fastest, most performant version of MongoDB we’ve ever built, designed for applications and agents that act on live data.
• Atlas Infinite is a new deployment option built for highly elastic workloads, allowing storage and compute to scale independently as demand changes.
• Atlas Agent Engine provides the memory, runtime, and governance needed to operate agents in production, on the same data foundation customers already trust and open to any model, framework, or cloud.

Before I share more about what we’re doing, it’s worth explaining the problem we’re trying to solve.

Why this moment demands more from the data layer

Almost every data company now describes itself as a platform for AI. The more interesting question is what happens to the data architecture when AI stops just answering questions and starts taking actions.

Consider a customer-service agent deciding whether to issue a refund, a commerce agent promising delivery of an item, or a financial agent taking action on an account. The agent needs to know what is true right now. A model can reason perfectly over the information it has and still make the wrong decision if that information is stale. Yesterday’s inventory, an old account balance, or an outdated order status can turn a technically sound inference into a bad action.

That makes freshness much more important. For many agentic applications, periodically copying operational data into another system isn’t enough. Agents need access to the business's live state, and increasingly they need to act on that same state.

But having current data doesn’t solve the entire problem. An enterprise may have enormous amounts of potentially relevant information. The agent has to identify the small subset that matters for the decision it is making. Better retrieval improves the context presented to the model, reduces irrelevant tokens, and lowers the likelihood that the model has to infer what it doesn’t know.

This is why we believe operational data and retrieval are converging. MongoDB is already the operational system of record for some of the world’s largest enterprises and fastest-growing AI companies. With native search and vector search, together with Voyage AI’s retrieval and embedding technology, developers can retrieve relevant context from the same live operational data their applications already use. You don't need to choose between operational data and AI context. Increasingly, they are the same thing.

Agents also create a very different workload. Traditional applications tend to have relatively understandable traffic patterns. Agentic systems can behave very differently. One user request may trigger a sequence of planning, retrieval, tool calls, database reads and writes, and additional reasoning. One agent may generate hundreds of operations. Thousands of agents may begin working concurrently.

As a result, the relationship between users and infrastructure consumption becomes far less predictable. Demand can be quiet one minute and enormous the next. Provision for the average and you risk falling over during a surge. Provision for the peak and you pay for infrastructure that sits idle much of the time.

We think the database architecture has to change with the workload. Compute should expand quickly when agents need it and contract when they don’t. Storage shouldn’t have to scale simply because compute does. And developers shouldn’t have to redesign their applications every time their workload crosses another threshold. That is the idea behind Atlas Infinite.

There is another architectural shift underway that I think is under appreciated. Useful agents are stateful.

They need to remember what has happened, retrieve what matters, understand who or what is taking an action, maintain context across interactions, and operate within explicit permissions and policies.

Today, developers often assemble that infrastructure themselves. They combine an operational database with a vector database, memory system, orchestration framework, model provider, and governance layer. Each additional system creates another integration point, another security boundary, and often another copy of the data.

That may be manageable for a prototype. It becomes much harder when agents take consequential actions inside a production enterprise.

We believe the state, memory, retrieval, and governance layer for agents should sit as close as possible to the operational data those agents act on. That is what Atlas Agent Engine is designed to provide.

Today's announcements build on MongoDB’s foundation.

What we’re announcing today

MongoDB 9.0

Agents operating on live data put enormous demands on the database underneath them.

MongoDB 9.0 is the fastest, most performant version of MongoDB we’ve ever built. Compared with MongoDB 8.0, it delivers up to 2x higher throughput on large instances, 35% faster reads, and 30% faster updates, significantly improving price-performance at scale.

We’ve also continued to raise the bar on security, observability, and developer productivity. Queryable Encryption now protects more types of data in more situations, while expanded private networking in Atlas makes it easier for organizations in highly regulated industries to run their most sensitive workloads on MongoDB.

The goal is straightforward. As applications become more autonomous and operate at greater scale, the operational database underneath them has to become faster, more secure, and more efficient.

Atlas Infinite

Agentic workloads can be extraordinarily bursty and Atlas Infinite is designed around that reality.

It represents the biggest architectural change we’ve made to Atlas, allowing storage and compute to scale independently so customers can scale each according to what their workload actually requires.

The result is an architecture designed to grow from prototype to petabyte scale without forcing developers to predict tomorrow’s infrastructure requirements today. Time to scale is more than 96% faster, and each shard can hold 10x more storage.

Just as importantly, customers don’t have to rewrite their applications to use it.

Atlas Infinite is a new cluster type inside the Atlas platform developers already know. Customers can choose it alongside Atlas Core depending on the characteristics of their workload.

Public preview begins on AWS, with full multi-cloud availability planned for general availability (GA) .

Atlas Agent Engine

The hardest part of building an agent is no longer the demo, it’s operating the agent reliably in production.

Atlas Agent Engine brings the stateful infrastructure agents need onto the same data platform where their operational data already lives.

Built-in memory and retrieval give agents access to relevant, real-time context while reducing unnecessary token consumption. A unified control plane provides identity and governance over agent actions. And because Agent Engine is open to any model, framework, or cloud, developers can choose the AI stack that best fits their application rather than being locked into a single provider.

This particularly matters for enterprises.

More than 70,000 customers and over 75% of the Fortune 100 already trust MongoDB with their operational data. Atlas Agent Engine gives them a way to build and operate agents on that same foundation rather than creating an entirely separate data architecture for AI.

The road ahead

We are still early in this transition. Models will get better. Context windows will get larger. Inference will get cheaper. Agents will become more capable and autonomous. But none of these advances eliminate the need for high-quality operational data. They make it more important.

As software moves from systems that record what people do, to systems that help people decide what to do, to systems that increasingly take action themselves, the data layer becomes part of the control plane for the business.

That is the opportunity we are building MongoDB for.

The name at the top of this post may be familiar, but MongoDB is much more than its leader. It’s the product and engineering teams who continue to deliver meaningful innovation, the go-to-market teams who obsess over customer outcomes, and the customers who trust MongoDB with their most sophisticated and strategic applications.

I’m grateful to all of them for what went into today. See you tomorrow at .local NYC.