Devin’s 2025 Performance Review is a post by the Cognition team, published on 14 November 2025, that assesses its autonomous coding agent Devin eighteen months after launch by borrowing the form of an annual review for a human engineer. The post states that Devin now works in engineering teams at thousands of companies and has merged hundreds of thousands of pull requests.
Cognition explains that it first tried to place Devin on a traditional engineering competency matrix and found this difficult: where human engineers tend to cluster around one level, Devin is senior-level at codebase understanding but junior at execution, and has unlimited capacity while struggling with soft skills. The post therefore summarizes strengths and weaknesses observed in real deployments instead, with examples and metrics that Cognition attributes to its customers. The performance and customer metrics in it are the vendor’s own report about its own product.
Key Points
- Cognition’s first strength pattern is “junior execution at infinite scale”: Devin does best on tasks with clear, upfront requirements and verifiable outcomes that would take a junior engineer four to eight hours, and unlike a person it can be run in parallel without limit.
- The post places vulnerability remediation, language and framework migrations, unit-test writing and small tickets in that category, describing them as critical but less creative work that frees human engineers for higher-impact projects.
- For migrations, the post says that once Devin is given instructions on how to update each repository, a fleet of Devins can execute the change across every repository in parallel; the post cites a bank’s proprietary ETL framework migration and a Java version upgrade as customer examples.
- Test generation is described as a division of labour: humans write a unit-testing playbook spanning a few hundred repositories, a fleet of Devins writes the tests, and code owners then check that all logic has been tested. Cognition reports customers’ test coverage typically rising from 50–60% to 80–90%.
- Cognition reports that over the preceding year Devin became four times faster at problem solving and twice as efficient in resource consumption, and that 67% of its pull requests are now merged, against 34% a year earlier.
- For brownfield feature work, the post says Devin can replicate and modify code where existing code provides clear patterns, and that Devin pushed about a third of the commits on Cognition’s own web app.
- The post limits Devin’s role in pull-request review to a first pass that catches obvious issues, stating that human review is still necessary because code quality is not straightforwardly verifiable.
- It describes data analysis and quality assurance as unexpectedly strong uses, with users asking Devin questions from Slack or asking it to build dashboards.
- The second strength pattern is “senior intelligence on demand”: Cognition says Devin has become much better at understanding large codebases — one driver, in its account, of the doubled merge rate — so it can document large codebases through DeepWiki and help engineers plan through a chat interface that explains systems with architecture diagrams, maps dependencies, flags breaking changes and recommends what should be done by humans versus AI.
- The first area for improvement is ambiguous requirements: the post says Devin cannot independently take an ambiguous coding project end-to-end using its own judgement as a senior engineer could, and gives visual design, where it needs specifics such as component structure, colour codes and spacing values, as an example.
- The second is scope change: Devin handles clear upfront scoping well but usually performs worse when told more after starting a task, which the post contrasts with a human junior who can be coached through iterative problem solving. It concludes that engineers carry more responsibility for scoping work up front and have to learn to “manage” Devin.
- The third is soft skills: Devin collaborates in Slack, Teams and Jira but cannot manage reports or stakeholders or deal with teammates’ emotions.
Context
The post is a vendor’s account of its own product and closes by inviting readers to talk to Cognition’s sales team; its customer outcomes are reported by Cognition rather than by the customers themselves. Its stated weaknesses and strengths rest on the same condition: clear, verifiable requirements are where the post says Devin succeeds, and ambiguous or shifting requirements are where it says Devin falls short.
For 2026, Cognition states it will continue to work on making Devin better at understanding real-world codebases and on using that context to collaborate with engineers on end-to-end software engineering work, and is investing in the user experience so that Devin is easier to direct in everyday development.