Cognition and AWS Sign Deal: Devin Cuts COBOL Modernization From 8 Months to 8 Days

Cognition and Amazon Web Services have signed a multiyear Strategic Collaboration Agreement to deploy the autonomous engineer Devin inside enterprise AWS environments. The companies' main argument is customer case studies. Mercedes-Benz ran more than 200,000 lines of COBOL through Devin, and the estimated modernization timeline fell from eight months to eight days. Another global automaker moved a 25,000-line COBOL process from a mainframe to AWS Lambda, at an estimated 73% lower cost.

Devin can already be bought in AWS Marketplace, and Agent Toolkit for AWS is available directly inside the agent. The two companies are also exploring deeper integrations to make it easier to embed Devin in customers' existing AWS environments. Joint customers named so far include Mercedes-Benz, Echo Global Logistics, and ActiveCampaign.

What work the agent takes on

Cognition is betting on tasks that engineering teams never get around to: migrations, framework upgrades, vulnerability remediation, and maintenance of legacy systems. Devin can study a codebase, draft a plan, write code, test the result, and fix the problems it finds. Humans decide what to build. They also review the output.

In modernization projects, the approach works like this: Devin analyzes the existing system, reconstructs its behavior, writes replacement code, and checks the new system against the old one's known outputs. For legacy code, this is a workable technique. Such systems often have no documentation, and their behavior is captured in their outputs. Those outputs become the reference standard for verification.

Key details of the announcement:

• Deal format: a multiyear SCA between Cognition and AWS.
• Purchase channel: AWS Marketplace, plus Agent Toolkit for AWS inside Devin.
• Mercedes-Benz case: more than 200,000 lines of COBOL, project estimate cut from 8 months to 8 days.
• Automaker case: 25,000 lines of COBOL, mainframe → AWS Lambda, estimated 73% cheaper.
• Parallelism: multiple simultaneous sessions for migrations, upgrades, and security fixes.

Isolation and control

For an enterprise buyer, the question usually comes down to where the code and data live. Devin is deployed in a single-tenant environment on AWS, and the customer picks the region. Data and organizational knowledge stay in a dedicated VPC, which helps meet internal policies and data localization requirements.

Each session starts in its own virtual machine, which is created for the task and deleted when it finishes. The audit log API records user actions and passes them to the monitoring and compliance systems the customer already runs. Platform teams apply the same regional, network, and audit restrictions to Devin that they apply to the rest of their AWS infrastructure.

Art Levy, Cognition's vice president of global partnerships, says engineers have been held back for years by legacy systems. According to him, Devin lets companies put autonomous agents alongside people and tackle postponed work. Anasuya Strasner of AWS calls it a way to speed up migrations and clear technical debt on infrastructure customers already trust. Christina Garcia, SVP of engineering at Echo Global Logistics, notes that her team uses Devin both to understand codebases through DeepWiki and to close gaps in older systems.

Where teams should start

If the backlog includes migrations with a clearly defined input and output, such as batch processes or calculation modules, start there. The agent has a reference standard to check against. Before a pilot, it helps to assemble a set of known inputs and the old system's expected results. How convincingly you can demonstrate equivalence depends on that set.

It is worth agreeing in advance with platform and security teams on how Devin's audit log will connect to current monitoring and which regions are acceptable. And keep in mind that the figures cited, eight days instead of eight months and 73% savings, are estimates from the companies' own materials. Before building a business case, verify them in your own pilot and account for the review time that remains with engineers.