Agentic pull requests show significantly lower merge rates
Linear B data indicates that agentic pull requests are merged at a rate of 33%, compared to 88% for human-written code. The disparity is attributed to the lack of human advocacy for agent-generated code.
Linear B’s 2024 AI benchmark data reveals a significant gap in the success rates of agentic versus human-written code. Across eight million pull requests tracked in the first half of the year, only 33% of agentic pull requests were merged in the most advanced organizations. In contrast, 88% of human-written pull requests were merged. The data suggests that agent-generated code often fails to reach production because it lacks a dedicated 'code owner.'
Human developers actively advocate for their code by seeking reviews and resolving conflicts, whereas agent-generated code frequently stalls without this human intervention. Thomas Dohmke, CEO of Entire and former CEO of GitHub, emphasized that every pull request, regardless of whether it is generated by a swarm of agents, requires a business owner who is ultimately responsible for the code reaching production.
This finding highlights a shift in engineering workflows, where the focus is moving from writing code to reviewing and governing the output of AI agents. To ensure code quality and completion, organizations must establish clear ownership structures. The data underscores that agentic code requires explicit human ownership to be successfully integrated into the development lifecycle, as agents do not naturally advocate for their own work in the review process.