Cognition builds its Devin agent into MongoDB's platform: migrations 4–5 times faster

Cognition and MongoDB have connected the Devin agent to the Application Modernization Platform (AMP), MongoDB's tool for moving legacy applications to Atlas. The product is called Devin for MongoDB Modernizations, and it is already available to customers of both companies. In early joint tests, tasks that used to take five to six hours now take just over an hour, roughly 4–5 times faster. The vendors' promise is ambitious: migrating decades of accumulated applications in months rather than years.

Where migrations actually get stuck

The wording of the announcement is telling: migrations to MongoDB rarely stall because of the database itself. The problem is the "wiring" around the old code: thousands of queries, stored procedures, and data access layers that have accumulated business logic over the years. A script can't untangle that. You need to understand why a particular query was written the way it was, and what will break if it is replaced.

This is the territory Cognition is aiming Devin at. The agent works across the entire codebase: it plans the changes and rewrites the business logic and the data access layer, handling each case individually. This is exactly the kind of work that typically eats up engineering capacity, as Cognition CEO Scott Wu says directly: even the most ambitious teams spend a significant share of their resources maintaining old systems.

How Devin and AMP split the work

The architecture rests on a division of responsibility between a non-deterministic component and a deterministic one. For data migration, the risk of hallucination is unacceptable, so AMP's proven tooling handles the transfer and verification of every record in MongoDB. The code, where flexibility is needed, goes to the agent. Devin connects directly to AMP's tools and orchestrates the process end to end, so the customer sees one coherent migration rather than a collection of scripts and manual handoffs between teams.

Key details from the announcement:

• Devin's role: planning, and rewriting business logic and data access layers.
• AMP's role: deterministic transfer and validation of records in MongoDB Atlas.
• Left to humans: designing the target data model and the cutover sequence.
• Early test results: from 5–6 hours to just over an hour per task.
• Availability: immediately for MongoDB and Cognition customers; documentation for connecting Devin to MongoDB has been published.

The line separating what humans decide from what the agent does is drawn sensibly. The data model for a document database is an architectural decision with consequences that last for years, and the order of the switchover, which depends on business risk, should not be delegated to an agent.

What we know about the results, and what we don't

The only published metric is "in early joint tests." The methodology has not been disclosed: it is unknown what counted as a task, which codebases were used for measurement, and how human review time was accounted for. The material appeared on the blog of Windsurf, which is affiliated with Cognition, so this is the vendor describing itself, not an independent assessment.

A more concrete voice in the announcement is Bilt. CTO Kosta Krauth says the company writes most of its software with Devin and deploys to production daily. Neighborhood search on bilt.com, the first thing a user sees, runs on MongoDB; with Devin, Bilt is moving its remaining search workloads onto the same database on timelines that would be impossible by hand. That is one customer and one scenario, but it shows how the combination is used in a real product.

MongoDB has its own motive. CEO Dev Ittycheria puts it this way: the value lies in modernizing as much of an enterprise's application estate as possible at the same time, rather than one application at a time. At the same time, the company is pushing the thesis of a modern data foundation for AI applications and agents that run on live operational data. Here, migration is a funnel into Atlas.

How to apply this in practice

If you have legacy applications on relational databases and a MongoDB migration sitting in your backlog, it makes sense to request a pilot on a single non-critical service. Measure not only code generation speed but also the full time to production, including review and data validation.

Fix the target data model and the cutover plan in advance: these are the decisions that stay on your side, and they determine how much the agent speeds up the work. And check how AMP validates records: trust in the result depends more on the completeness of that check than on any figures in the announcement.