# How Forward Deployed Engineering is Done at Cognition Jia Wu, Cognition | AI Engineer World's Fair 2026 | 17:38 Source: https://www.youtube.com/watch?v=RVxym6mmIns Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/how-forward-deployed-engineering-is-done-at-cognition Published: 2026-07-28 Tags: agents, coding-agents, deployment, enterprise, product-strategy ## TL;DR - Forward-deployed engineers connect Cognition's product capabilities to the customer's highest-value engineering problems. - The job measures delivered outcomes such as delivery timelines, productive engineering capacity, and pull requests instead of token usage. - FDEs bring recurring customer problems back into the product roadmap while staying responsible for correctness and customer success. ## Summary Jia Wu explains how Cognition deploys Devin inside large enterprise customers. The FDE team starts by understanding the customer's software work, including backlogs, migrations, testing, alerts, reviews, and maintenance. It then points the agent at work with business value instead of running sessions simply to consume tokens. Wu describes deployments through several outcome measures: engineering capacity, delivery timelines, and pull requests. One case study produced about 150% additional headcount over three months, while another showed an 82% reduction in delivery timelines after Devin was fully activated. The role also feeds customer problems back into Cognition's roadmap, helping the product address recurring enterprise challenges. Wu describes FDEs as people with broad business, customer, process, and technical skills, with a deep specialty in at least one area. He closes by tying the work to correctness, direct customer commitment, and shipping difficult projects without leaving work undone. ## Key ideas ### Cognition's internal use of Devin produced a step change in engineering output [01:57](https://www.youtube.com/watch?v=RVxym6mmIns&t=117s) Wu says Cognition used its own agent during a period when the company may have been behind on hiring and still shipped almost an order of magnitude more good-quality pull requests across the organization over six months. He calls this a "step function increase" in engineering leverage. Customer adoption grows differently, he says, with companies deploying the agent across multiple use cases and scenarios in a parabolic pattern. The difference comes from broader organizational use rather than a single developer using a tool. ### Forward-deployed engineers increase the overlap between the product and customer problems [02:29](https://www.youtube.com/watch?v=RVxym6mmIns&t=149s) Wu frames the work with two circles: the product's domain and the customer's problems. Their overlap is where the company can deliver value. Cognition's forward-deployed motion tries to increase that overlap by deeply understanding the customer's problem space. In software engineering, that includes building features, reviewing code, deploying it, testing it, and maintaining it. Wu says writing code is usually only about 20% of the problem. The rest involves getting software safely through the development lifecycle and keeping it working across the enterprise. ### Agents need a clear target or they mainly consume tokens [05:14](https://www.youtube.com/watch?v=RVxym6mmIns&t=314s) Wu says deploying agents without a specific direction amounts to "token maxing": spending tokens without producing tangible outcomes. FDEs spend time with customers to identify strategic initiatives that matter most to the business, then map Devin's capabilities onto that work. His example of the job includes four or five hours of customer calls followed by four or five hours of hands-on keyboard work. The goal is also to automate the deployment itself, so agents can respond to alerts and events instead of requiring someone to trigger every run manually. ### Customer deployments feed recurring engineering problems back into Cognition's roadmap [07:07](https://www.youtube.com/watch?v=RVxym6mmIns&t=427s) Enterprise challenges often arrive in similar shapes, according to Wu. FDEs hear about them in the field and bring that context back to product. They ask whether a problem is common across enterprises or specific to one user, and whether a workaround, hack, or bug should become a product feature. This feedback helps reduce uncertainty about what to build. Wu describes FDEs as the bridge between products and problems, with each deployment improving the next one. Solving a customer's immediate issue is only part of the work; the product must improve too. ### FDEs need broad customer and business skills plus a deep specialty [08:34](https://www.youtube.com/watch?v=RVxym6mmIns&t=514s) Wu describes the role as flexible and T-shaped. FDEs need range across people skills, business, process, customer work, and technology, along with a deep spike in at least one area. Cognition hires people from product management, founding, and software engineering backgrounds. Product managers may bring a strong sense of how products fit together, while technical specialists can bring expertise that is difficult to learn during a deployment. Wu says a good customer engineer can connect customer problems to the roadmap, while a great one keeps asking why the problem matters to the business. ### Cognition measures deployment value through outcomes rather than agent activity [10:25](https://www.youtube.com/watch?v=RVxym6mmIns&t=625s) Wu contrasts an earlier period when deployed engineers were encouraged to maximize token usage with the current focus on delivery. Enterprise customers want to know whether the solution creates real value or simply burns tokens. Cognition tracks agent sessions, engineering hours generated, and productive engineering hours. Wu gives a case study in which the team embedded with a customer for three months and delivered about 150% additional headcount. He says the team then checks whether the activity is meaningful by looking at delivery timelines and the number of pull requests shipped. ### One fully activated deployment reportedly cut delivery timelines by about 82% [12:56](https://www.youtube.com/watch?v=RVxym6mmIns&t=776s) Wu says Cognition compares metrics from before Devin arrives with metrics after it is fully activated inside the customer environment. In one set of delivery work, the reported reduction was about 82%. He also says Cognition delivered almost twice as many pull requests as engineers using single-point tools before an agent harness like Devin was introduced. He presents these as separate proof points: added engineering capacity, shorter time to value, and more shipped pull requests. The examples are anonymized, alongside public customer examples. ### The role stays accountable for correctness and customer success [15:53](https://www.youtube.com/watch?v=RVxym6mmIns&t=953s) Wu says Cognition's FDEs carry company principles into customer work. The team can spend months embedded with a customer, including placing someone in Brazil for ten months to live near one customer. He describes correctness as more important than protecting ego: engineering practices and product issues should be raised and sent back to the product team. FDEs may sit between sales and post-sales, but Wu says everyone is part of go-to-market because the target is customer success at all costs. The work is defined by shipping and by acting quickly because every second counts. ## Notable quotes - "The problem isn't writing code faster, that's usually only 20% of the problem." (04:27) - "If we are taking it union of the problems that the customers have and the products that we are building, solving the problem only shifts part of the Venn diagram." (10:38) - "You can make engineers like 10x faster. That's fine. That's still valuable. But can you make an organization 10x faster, including every single person that might be technical or non-technical across the company?" (11:32) - "At the end of the day, we are all in the same boat. We're on the same mission and we're just shipping." (16:40) ## Tools & references mentioned - Cognition - Devin - Windsurf - Cursor - Nubank - Cockroach Labs - Built - COBOL - JCL ## Who should watch - You are building or deploying coding agents inside a large, regulated enterprise and need a way to measure delivered value. - Your customer-facing engineering role includes translating messy production problems into product and roadmap decisions. - You want to understand what Cognition expects from a forward-deployed engineer beyond writing code or running agent sessions.