Cognition has invested nearly three years building Devin, an autonomous coding system designed to accept tickets and produce pull requests. The company's latest move, announced Tuesday with the introduction of Devin Bots, shifts focus toward orchestration—deploying persistent agents capable of managing multiple Devin sessions and owning engineering outcomes across days or weeks.
Coordination becomes the product
CEO Scott Wu contends that even marginal reductions in human effort could yield substantial productivity improvements. In a post on X, he posed a question about developer capacity if coding agents advanced from 98% to 99% task coverage, noting that Cognition observed significant variance in outcomes across beta participants and internal groups.
The percentages are hypothetical, but Wu is making the case that coordination, not just coding, is the next opportunity for automation.
From tickets to targets
Silas Alberti, a founding team member at Cognition, characterized the shift in an X post as moving away from discrete task assignment toward agents pursuing continuous objectives. According to Alberti, the team has already demonstrated the capability to compress research timelines from weeks into days using Devin Bots.
Cognition illustrates the approach with a concrete example: maintaining feature parity between Android and iOS versions of an application. When iOS receives new functionality, a Devin Bot can initiate and coordinate corresponding Android development without requiring an engineer to manually generate tickets for each feature addition.
This direction extends an emerging pattern in which coding agents operate as collaborative team members. Devin Bots advances this model by positioning an agent as the coordinator of work distribution and execution.
An Android app, barely supervised
When Cognition decided to develop an Android version of its own platform, the company lacked dedicated Android engineering resources. Instead, it tasked a Devin Bot with the project on September 28 and allowed the agent to manage the technical work independently.
During the subsequent 44-day period, Cognition reports the Bot integrated 1,134 pull requests, generated 344,651 lines of Kotlin code, and resolved 1,217 tickets. The Bot also executed 2,841 screenshot tests. The supervising engineer contributed 183 messages across this timeframe, with 22 days passing without any communication from the human team member.
Cognition's account describes how the Bot created tickets and routed them to individual Devin sessions, then consolidated changes after they satisfied continuous integration and Devin Review requirements. Daily Slack notifications with screen recordings allowed the team to monitor progress. The resulting Android application entered open testing via Google Play, though additional development remains before achieving feature completeness with the iOS counterpart.
In another deployment, Cognition's Layout Bot scans the company's web application for visual defects every three hours, recompiles the latest production build, and validates the interface across varying load scenarios. Engineers approved 159 of its corrections over three weeks, achieving a 70% reduction in layout latency, per Cognition's report.
Cognition says the studio went from producing around 50 pull requests a week to 1,900.
A mobile game development studio operates roughly 20 bots under the coordination of an agent called Super Jenny. Super Jenny surfaces decisions to a human operator four times daily. Cognition reports the studio's pull request output increased from approximately 50 per week to 1,900.
Where oversight runs thin
When multiple coding agents work in parallel on the same codebase, conflicting dependencies or bugs affecting other agents' contributions become possible. Cognition has not detailed how Devin Bots identify and resolve such conflicts.
Tasks requiring approval go to a "Needs You" inbox for an engineer to review.
Engineers can establish guardrails by protecting specific CI checks and imposing compute spending caps before delegating work to a Bot. Code requiring human judgment routes to a "Needs You" inbox. However, Cognition has not revealed the decision-making process by which the supervising Bot determines code readiness for merging or escalates repeated failures to human teams.
Test authorship by agents presents additional validation challenges. Agent-authored pull requests are already straining traditional code review workflows, and some organizations are establishing separate verification mechanisms to catch defects that the original coding agent might overlook.
Cognition has not disclosed the volume of engineering time spent reviewing Bot-generated pull requests or remediating failed implementations. The New Stack requested additional detail on how Devin Bots coordinate parallel agents, assess work quality, and determine escalation thresholds.