You have 1 article left to read this month before you need to register a free LeadDev.com account.
Estimated reading time: 3 minutes
The rise of agentic AI coding tools is causing a rethink of approaches to massive monorepos.
The latest generation of large language models (LLMs), and coding assistants like Claude Code and Codex is significantly increasing the volume of code organizations need to handle.
Boris Cherny, the creator of Claude Code, has written extensively in recent weeks about how he’s processing around 250 pull requests a month – roughly eight or nine pull requests a day. (Some studies have the average at 12 a month.) He even admitted he barely, if ever, writes a line of code these days, saying that 100% of his recent contributions were written by Claude Code rather than typed by him directly.
This surge in code volume threatens to overwhelm traditional version control tools, particularly when managing massive monolithic repositories (monorepos).
This shift has started the rumour mill, with speculation that big tech firms are having to refactor projects quickly to avoid succumbing to an onslaught of code commits of their own making.
Your inbox, upgraded.
Receive weekly engineering insights to level up your leadership approach.
“The new goal of most AI labs today is to deploy 1,000 coding agents for a team of 10 supervising engineers,” believes Son Luong Ngoc, solution engineer at BuildBuddy.
Jake Cooper, the founder of Railway, and Geoffrey Huntley – inventor of the Clawdbot Ralph sequencing that allows agentic AI to run autonomously overnight in repeated performance improving loops –have both suggested that it’s important for companies to start reconfiguring their systems now. (Cooper did not respond to a request for comment.)
“Monorepos still work well, but you need to be good at detecting what dependencies have changed, and therefore what part of the codebase must be rebuilt and retested,” says Liz Fong-Jones, technical fellow at Honeycomb.
“This is a problem Google already solved 15 years ago,” she adds, pointing to Google rolling out their Blaze – now Bazel – system which precisely detects changed dependencies and only rebuilds and tests code paths that are affected. “I have no doubts about their ability to do this, but smaller startups? Maybe not.”
A torrent of code
Early data also suggests just how quickly the torrent can overwhelm existing defenses.
Researchers from Drexel University and Missouri University of Science and Technology analyzed more than 33,000 agent-authored pull requests across GitHub, examining the output of five major agents including OpenAI Codex, GitHub Copilot, and Devin.
The success rates of the agentic AI tools varied wildly: OpenAI Codex achieved an 82.6% merge rate, far outstripping GitHub Copilot, where only 43% of its contributions merged, suggesting that simply turning on the “agent tap” doesn’t guarantee usable code.
Failed agentic PRs tended to be significantly larger and more invasive than successful ones, often modifying more files and lines of code. They were also far more likely to fail continuous integration (CI) checks, with the study noting that each additional failed CI check decreased the odds of a merge by roughly 15%.
“LLMs definitely are good at writing voluminous and exhaustive tests, but now the CI infrastructure is going to creak under the amount of parallel tests being requested by so many different pull requests and the number of tests each PR needs to run,” says Fong-Jones, who wasn’t involved in the study.
More like this
Revamping the code factory
The researchers concluded current agents struggle with “coordination and alignment,” frequently submitting unwanted features or ignoring explicit reviewer instructions. Without better tooling they risk becoming a DDoS attack on your development pipeline, rather than a productivity boost.
Fong-Jones suggests that an important issue is helping agents find the correct part of the codebase to work in. “With a monorepo, you certainly have hierarchical directory structures which help, but AGENTS.md files need to be sprinkled all over to steer the agents,” she says. “Otherwise they’ll wind up with either insufficient knowledge of local patterns,” she says, or could also use so much of the AI’s context window with codebase that they lose performance when it comes to actually making changes.
“This will create a massive volume of new code and calls for a revamp of the traditional ‘code factory’,” says Luong, who was also not involved in the study.
London • June 2 & 3, 2026
LDX3 London agenda is live! 🎉
As a result, it’s only natural that tech companies are starting to adapt to that new reality. “The processes to review code, test code, build code, and deploy the code are all changing rapidly to accommodate for this inevitable future,” he explains.
If an organization is wary of a mass reconfiguring, at a start, it could include better build tooling, AGENTS.md files, and smarter dependency graphs.
