AWS details how monday.com runs production AI coding agents at scale on Amazon Bedrock

AWS has published a detailed account of how monday.com runs "AI Teammates" — agentic AI built on Amazon Bedrock — in production across a decade-old codebase, reporting that nine in ten of its engineering "Builders" now use AI coding tools every month and per-engineer pull-request throughput is up more than half.

What's new

AWS writes: "AI Teammates are agentic AI on Amazon Bedrock, and few engineering organizations run them in production at the scale that monday.com does. Nine in ten Builders use AI coding tools every month, up from roughly half a year ago. Per-engineer PR throughput is up by more than half. Every figure in this post comes from monday's own internal production data." The post frames monday's rollout as three levels of AI-engineering maturity: L1, where engineers use AI as a pair programmer (Cursor for fast reflexive work, Claude Code for heavier lifts, with adoption nearly doubling year over year); L2, where teams build reusable skills and sub-agents for repeated work with engineers still in the driver's seat — where AWS says most of monday's engineering runs today, and where per-developer PR throughput stepped up by more than half; and L3, fully agentic delivery where agents own work end-to-end while engineers orchestrate. AWS is explicit that this is not a greenfield deployment: monday.com is described as a decade-old code base serving millions of paying users across hundreds of microservices and microfrontends, where every agent-opened PR goes into a system expected to keep working through the next deploy.

Context

The post continues a run of AWS enterprise case studies on agentic AI in production, following its June 30 post on safely releasing frontier models to customers and its June 10 post on frontier engineering teams reinventing AI-native development. Where those posts made general claims about productivity gains, this one is built entirely around one named customer's internal metrics.

Why it matters

Concrete, named-customer production data on coding-agent ROI — rather than vendor benchmarks or anonymized survey results — is still uncommon, and monday's L1/L2/L3 framing gives other engineering organizations a rubric to locate their own AI adoption maturity. The specific detail that most of the reported PR-throughput gain came from L2 (reusable skills and sub-agents with engineers still directing the work, not full agent autonomy) is a useful data point for the broader industry debate over how much of today's measurable agentic-coding value actually requires giving up human control versus simply automating well-scoped subtasks.