On July 22, 2026, AWS published a detailed account of how monday.com put AI agents into production on Amazon Bedrock. The post reports two striking outcomes: nine in ten Builders now use AI coding tools every month, up from roughly half six months prior, and per‑engineer pull request throughput is up by more than half. AWS says every figure comes from monday.com’s internal production data. The system, branded as monday.com AI Teammates, is built to execute real work inside a decade‑old codebase without breaking the safety rails.
Inside monday.com AI Teammates on Bedrock
According to the AWS blog by Claudio Mazzoni, Erez Drutin, Moran Zilberstein, Netanel Abergel, and Ofek Dayan, monday.com’s approach mixes Bedrock’s model orchestration with well‑known AWS primitives: Amazon Elastic Kubernetes Service (EKS), Amazon Elastic File System (EFS), Amazon Relational Database Service (RDS), Amazon Simple Notification Service (SNS), and Amazon Simple Queue Service (SQS). The post frames AI Teammates as agentic workers that participate in the company’s software lifecycle, push code, and move changes toward merge using confidence scores.
The stack choice matters. Bedrock provides a managed entry point for foundation models, while EKS gives monday.com a way to run agent services as containers they already understand. EFS can support shared artifacts across pods, and RDS can hold state that must persist across tasks. SNS and SQS are the connective tissue for events and backpressure. Together, the parts point to a design built for scale and failure isolation, even as the agents touch production workflows.
What sets this apart is not just another assistant that drafts code. The post outlines a confidence‑scored merge play designed to close the gap to full autonomy. In plain terms: the system knows when it should ask for help, when it can proceed, and when to stop. That gating is how monday.com puts AI Teammates inside the development flow without handing over the keys.
How production Bedrock agents were wired
The architecture choices signal a bias for components teams already operate. Amazon Bedrock handles model access and guardrails. EKS runs agent microservices that scale horizontally and roll back cleanly. Messaging spans SQS for work queues and SNS for fan‑out, which keeps the agents reactive without tight coupling. That pattern is familiar to ops teams, and it gives product owners a way to reason about bottlenecks and failure modes.
The authors emphasize that the codebase is about a decade old. That detail matters more than the shiny tools. Most enterprises are not greenfield. Getting agentic systems to work against battle‑scarred services demands retrofits, common interfaces, and predictable rollback. The described setup implies human‑in‑the‑loop gates, audit trails in a central store, and queues that can drain safely if something goes sideways. In practice, that’s how a company can let agents propose and advance changes while holding a line on risk.
There’s a cultural thread too. The AWS post claims near‑universal monthly use of coding tools across Builders, with a clear lift in per‑engineer PR throughput. Those are the adoption and velocity signals leaders ask for before widening access. The fact that monday.com AI Teammates live on the same backbone that runs the rest of their stack likely lowered the friction to try, measure, and scale.
What the adoption and throughput gains say
Two numbers jump out: 90% monthly usage and a more‑than‑50% rise in PR throughput. Both come from monday.com’s internal telemetry, per the AWS blog, so they are best read as directional. Even so, the pairing tells a story. Adoption without throughput can be a novelty effect. Throughput without adoption can be heroics by a few power users. Seeing both move together suggests a system that fits into existing workflows rather than fighting them.
It also hints at where the gains likely come from. Drafting code is the obvious starter. The AWS post points to a broader loop: tickets, changes, tests, and then merges guarded by a confidence score. That last step matters because it reduces the management overhead of triaging agent output. Engineers spend less time playing goalie and more time reviewing changes that have already cleared basic checks.
Put another way, the value appears in the handoffs. SQS work queues keep tasks flowing. SNS can wake the right services at the right moment. RDS anchors state that shouldn’t vanish mid‑flight. Bedrock supplies the model layer without asking teams to run their own inference stack. The combination helps the agents stay busy and measurable, and it gives leaders the auditability they need to justify broader rollouts.
The path to autonomy is gated by confidence
The most telling idea in the AWS post is the confidence‑scored merge. Full autonomy in software delivery isn’t a single switch. It’s a series of gates where the bar rises as evidence accumulates. A confidence score becomes a contract: if the agent’s work hits a threshold that past data says is safe, it moves forward. Miss the bar, and the agent escalates to a human or stops. That’s a practical way to climb toward automation without betting the codebase on a model’s best guess.
This framing also explains why monday.com parked AI Teammates on Bedrock and inside their standard runtime. You can only trust a score you can measure, and you can only improve what you can ship often. Bedrock abstracts model churn. EKS lets teams deploy small changes to the agent logic with the same pipelines they use elsewhere. Over time, that creates a data loop where the score gets sharper, the gates get smarter, and more work routes past manual checks.
For organizations staring at older systems, that’s the real takeaway. Don’t chase exotic new stacks. Place agents where you already have strong observability and safe deploys. Use queues and pub/sub to keep the blast radius small. Start with human review, then replace reviews with scores only when your data says it’s safe.
Why this matters beyond monday.com
The AWS write‑up reads like a blueprint many teams could follow. It shows how to wire Bedrock agents into production without waiting for a wholesale rewrite. It ties adoption to clear outcomes that executives track. And it treats autonomy as a spectrum managed by confidence and gates, not a leap of faith.
There are limits. The numbers are self‑reported by monday.com, as AWS notes, and practices that fit one company won’t map perfectly to another. But the architecture is full of defaults that enterprises already trust. That lowers the cost of trying, and it raises the odds that early wins translate into sustained use.
One more implication: the hiring market. If 90% of Builders at a large SaaS company are using AI coding tools, engineers entering the field will expect similar support elsewhere. Shops that lack a plan for agents in their pipelines may struggle to keep pace, even if their codebase is older. The monday.com AI Teammates story suggests they don’t need to wait.
The next questions are clear. How high can the confidence threshold rise before merges become a formality? Where does the agent stop—code, tests, rollout plans, or incident response? The answers will vary, but the route runs through the same terrain: Bedrock for model access, queues for flow control, and a runtime teams already know how to run. That’s how monday.com AI Teammates moved from pilot to production, and it’s a map others can use.