Cognition increased merged PR throughput 40x while its engineering headcount grew 2.7x over the same year.
2
Cognition runs development through small pods, Slack, local service environments, and agents that handle work from bug report through testing and maintenance.
3
Walden Yan argues that teams should give agents measurable jobs, such as improving latency or maintaining an Android app, instead of sending them isolated coding tasks.
Summary
Walden Yan explains how Cognition changed its engineering process around Devin. Merged PR throughput grew 40x in a year while engineering headcount grew 2.7x. The process starts with self-driving PRs, where an agent can take a bug report through implementation, testing, review, merge, monitoring, and maintenance, with humans handling escalations. Cognition reorganized engineers into small pods with responsibility for customer reports, QA, product decisions, and their own work queues. Much of the workflow runs from Slack. To make this possible, every service and mocked third-party dependency runs locally in Docker Compose inside disposable virtual machines that are rebuilt hourly. Yan then describes agents that inspect development data and improve the software factory itself. His broader proposal is to give agents measurable jobs, such as reducing P99 latency, tuning inference on new GPUs, or maintaining Android parity with an iOS app. The examples show a shift from agents completing assigned tasks to agents operating persistent systems with limited human input.
Cognition increased merged PR throughput far faster than engineering headcount
Walden Yan opens with a one-year comparison of Cognition's engineering output. Merged PRs grew 40x, while engineering headcount grew 2.7x. He says each engineer is now more than 10x more productive than before, and he considers the PRs at least as high quality as they were a year earlier. The talk focuses on the changes behind that result, including changes to the codebase and to the way engineering teams work. Yan presents these as practical changes for teams that want to build a software factory around agents.
Cognition makes the default development flow self-driving
Yan describes a software factory where a PR can move from an initial bug report to final merge, monitoring, and maintenance without a human involved. Human judgment still enters through escalations, but the default process is run end to end by Devin. This changes what engineers do. Picking tickets from a Linear board and sending them to an agent is no longer enough, because anything that can be described and implemented can be sent to Devin. Engineers need to take responsibility for the surrounding product and development system.
Small pods own customer problems from report through QA
Cognition broke larger teams into smaller pods, often with two or three engineers. Each pod gets swarms of agents and more vertical responsibility. Engineers listen to customer bug reports, perform QA, handle product management, and manage their own Linear boards. They then send implementation work to agents. The engineer's job covers the full problem rather than only writing code from a preselected ticket. Yan describes this as a fundamental change in how engineers work.
Slack becomes the front end for most of the development lifecycle
A customer can send a bug report and video in Slack, which Cognition forwards to an internal channel while tagging Devin. Devin produces a PR, then sends back a video showing the fix and its evidence. Yan says this lets him handle most development without opening an IDE or GitHub. The target is a workflow based on evidence, videos, and proof of a working change. The coding environment is still important underneath, but the people directing the work can stay in Slack.
Local disposable environments let agents test without live services
Cognition moved every service to run locally with Docker Compose and created local mocks for its third-party services. The complete setup runs inside a virtual machine that agents can access without touching live services. The company rebuilds these VMs every hour from the latest code. If a build fails, Devin is the first responder and fixes the VM snapshots. Each new Slack request therefore starts an agent on a fresh machine and a fresh version of the codebase. Devin also maintains its own testing skills as the codebase changes.
Agents inspect the development process and improve the factory
Yan says the activity around a cloud agent is already tracked across the codebase, including lines of code, PRs, decisions, and token use. Cognition runs agents over this data to find wasted time, identify work that could become scripts, and add tests. Devin posts daily suggestions such as changing the Node setup to reduce CI time or changing a lint script to catch more bugs before AI code review. These agents are not only producing application changes. They are also changing the system that produces those changes.
Yan distinguishes between giving an agent a task and giving it a job. For example, an engineer could ask an agent to improve application P99 latency. The agent would add telemetry, establish the current value, create alerts, break down the work, send reports in Slack, and provide daily updates. Yan asks engineers to turn anything they care about into a measurable number, including click latency, layout shifts, and onboarding retention. Once a metric exists, an agent can work on improving it over a long period without repeated prompting.
Persistent agents can tune hardware and maintain an app
Cognition used an agent to improve inference throughput on Nvidia's new Vera Rubin GPUs, whose kernels were not ready when the hardware arrived. The agent received a permanent set of GPUs and a goal for throughput, then tried different strategies over about a week. The best strategies went to a human engineer for final approval, and Yan says eight or nine useful changes came from this process. In another example, Devin maintained an Android app against the existing iOS app over 44 days, monitoring feedback, running live QA, and merging as many as 40 or 50 PRs on some days with little human messaging.
"We don't even want you to be prompting your agents most of the time. We want you to just be giving your agent goals that they can work toward over long durations of time."08:52
Who should watch
You are leading an engineering team that has adopted coding agents but still routes work through tickets, IDEs, and manual QA.
Your services are difficult for agents to test because they depend on staging systems, live databases, or third-party services.
You want to give agents ownership of a measurable engineering outcome, such as latency, inference throughput, or ongoing product parity.