monday.com Puts AI Agents Into Production on Amazon Bedrock: 50%+ More PRs per Engineer

monday.com has moved its agentic AI Teammates into production on Amazon Bedrock and disclosed internal metrics on how they perform: nine in ten of the company's developers use AI tools to write code every month, and pull request output per engineer has risen by more than 1.5x.

What numbers did monday.com report?

monday.com draws these metrics from its own production analytics, and all of them point to broad adoption of AI in development rather than a pilot. Six months ago, roughly half of its engineers used AI coding tools regularly. Now it is 9 in 10.

• Nine in ten developers ("Builders") use AI coding tools every month, up from about half six months ago
• PR output per engineer has grown by more than 50%
• The agents run on Amazon Bedrock, the AWS managed service for running and orchestrating large language models
• All figures come from monday.com's internal production data, not from a survey or a marketing report

According to the AWS Machine Learning Blog, monday.com is one of the few engineering organizations running agentic AI in production at this scale rather than as a demo.

How did the agents fit into a 10-year-old codebase?

The main engineering challenge was not the agents themselves but fitting them into a codebase roughly ten years old. Projects like this accumulate legacy patterns, nonstandard conventions and glue code that trip up general-purpose AI tools.

monday.com describes the work as "retrofits": changes that made the existing architecture usable by agents. These include context about the codebase, guardrails and integrations. Without them, the model produces code that is formally correct but useless in this particular project. The company credits these retrofits, not the choice of a specific model, for making the setup work at scale.

What is a confidence-scored merge?

A confidence-scored merge is the technique monday.com uses to close the gap to fully autonomous agents. The system assigns a confidence score to each proposed change and uses that score to decide whether to merge the code automatically or send it to a human for review.

The approach trades speed against control. Routine, predictable changes with high confidence go through without human involvement, while anything contentious stays under manual control. This is the last hurdle separating an assistant that writes code under supervision from an agent trusted to merge on its own.

Every figure in this piece comes from monday's own internal production data, according to the AWS

Machine Learning Blog.

What this means

The monday.com case is a rare example of agentic AI in development being measured with production metrics rather than promises: share of users, PR output, share of automatic merges. For teams with large legacy projects, the takeaway is simple: the bottleneck is not the model but the scaffolding around it, meaning context, guardrails and auto-merge rules.

FAQ

What are AI Teammates?

AI Teammates is monday.com's agentic AI, built on Amazon Bedrock. Unlike a suggestion-style assistant, these agents carry out development tasks and take changes through to a pull request, and some of those changes are merged automatically based on a confidence score.

What platform do monday.com's AI agents run on?

monday.com's agents run on Amazon Bedrock, the AWS managed service for running and orchestrating large language models. The company uses it to run production agents across its entire engineering organization.