Cognition announced more than $2 billion in Series E funding at a $48 billion valuation on September 8, 2026. One week later, it announced a multi-year collaboration with AWS to expand enterprise deployment of Devin. The practical enterprise story is how autonomous engineering fits into existing infrastructure and change controls.
The funding announcement describes the company’s growth strategy; the September 15 AWS collaboration describes purchasing and deployment options. Neither investment nor distribution is evidence that an agent’s code can bypass review.
What is available and what remains a partnership plan
Cognition says joint customers can purchase Devin through AWS Marketplace and use the Agent Toolkit for AWS. It also describes deeper engineering integrations as an area the companies are exploring. That wording matters: an existing purchasing route and a future integration are different procurement milestones.
A Marketplace listing can simplify buying through an established cloud relationship. It does not replace evaluation of permissions, contractual terms, or the product’s fit for a team’s repositories. Platform owners should verify the exact features offered in the contracted deployment rather than assume every public announcement applies to every edition.
The deployment boundary described by Cognition
The company describes a single-tenant AWS environment with a selected region and dedicated VPC. Sessions run in task-specific virtual machines, and audit-log APIs can feed customer monitoring systems. These are relevant controls for organizations that want agent activity to fit within existing cloud governance.
A network boundary does not make every action inside it safe. Review which systems the agent can reach, how credentials are provisioned, and who can approve changes. A coding session that needs to inspect application logs may not need privileges to rotate production credentials or alter cloud policy.
Choose work that has an observable definition of done
Migration and maintenance backlogs are natural evaluation candidates because they can be scoped into concrete tasks. A dependency upgrade can be checked against build outputs, affected tests, and runtime behavior. A legacy migration needs an existing baseline and explicit behavior comparisons.
Do not measure progress only by commits or lines changed. Track accepted changes, rework, reviewer time, and defects found after merge. Include tasks the agent cannot complete or should decline. Those results reveal whether delegation increases useful capacity or simply moves effort into the review queue.
What the funding changes for buyers
The round may support product development and enterprise capacity, but valuation is not an operating guarantee. Buyers still need support commitments, export options, incident handling, and a plan for recovering unfinished tasks. Long-running engineering work should leave enough evidence for a human to resume it.
Nerova’s editorial recommendation is to start with a restricted repository and a small group of clearly verifiable tasks. Expand privileges only when the execution and review process justify them. The AWS agreement makes deployment an easier conversation; reliability remains a property of the complete engineering workflow.