monday.com runs 100s of AI agents in production on AWS Bedrock

monday.com, a decade-old SaaS platform serving millions of users with hundreds of microfrontends, microservices, and hundreds of engineers (called Builders—a mix of Engineers, PMs, Analysts, and Product designers), has deployed AI agents at production scale on Amazon Bedrock. The company measures nine in ten engineers now using AI coding tools every month, up from roughly half a year ago, and per-engineer PR throughput has increased by more than half. These figures come directly from monday's internal production data.

The company frames its journey in three levels of AI engineering. L1 is the assistant tier, where engineers use AI as a pair programmer (via tools like Cursor and Claude Code), with adoption nearly doubling year over year. L2 is the skills and sub-agents tier, where teams build reusable agents for repeated work with engineers in the driver's seat; this is where most of monday runs today and where the per-developer PR throughput stepped up by more than half. L3 is multi-agent, fully agentic delivery where agents own work end-to-end while engineers orchestrate. At the heart of this work is Sphera, monday's internal agent system. Rather than presenting a job queue, Sphera shows a Teams page of humans and agents, each with a profile, manager, scope, and performance score. Atlas, the central agent in the post, has the role of Software Engineer, with the job of picking up tickets, writing PRs, and shipping features—the same backlog as any human teammate. This schema is not decorative; it means agents have stable identities that flow through Slack, GitHub, and monday, so humans tag, assign, code-review, and deactivate them like any other teammate.

The architecture uses seven core AWS services: Amazon SNS fans out external triggers (Slack mentions, monday item assignments, GitHub PR review requests) to per-team SQS queues. The monday Builders CoWORK, a set of SQS consumers on EKS, resolves which agent owns each message and routes it to the right agent runner pod. All three inbox types (Slack, monday item, GitHub PR review) hit the same agent session with the same memory and workspace on disk. The Claude Agent SDK serves as the runtime, wrapped by the monday-agent-sdk for provider neutrality (routing calls to Amazon Bedrock), cold-start optimization (warm caches so the first model call typically goes out in under a second), and a custom harness that encodes monday's opinions on evaluation, plugin composition, and integration with Slack, monday, and GitHub. State is split across services: live state (current task, run cursor, distributed lock, heartbeat, message log) moves to Amazon ElastiCache for sub-millisecond reads with self-expiring keys; sessions and cross-session memory live in Amazon Elastic File System as directories containing checked-out repos, encrypted secrets, and chronological logs. Atlas's daily diary (e.g., diary/2026-04-21.md) is a plain markdown file read at session start to resume prior work; for example, one entry notes that the accounts-api team files new error codes in a shared registry before the PR opens, a lesson learned from a prior reversion. Durable records (final transcripts, snapshots, artifacts, evals) go to Amazon S3, keyed by session ID. Amazon Bedrock also provides Application Inference Profiles for cost tracking, one audit trail for all model calls, cross-region failover, and AWS PrivateLink endpoints to keep traffic inside the VPC.

Five retrofits made agents viable inside a decade-old codebase. First, evals before model upgrades: early PRs broke once volume picked up, so monday added deterministic metrics (PRs merged, revert rate, zero-touch merge rate) per agent and LLM-scored evals across five dimensions per PR (Intent & Decision, Execution & Artifact, Completeness & Usefulness, Instruction & Boundary, Efficiency). Second, memory as a file rather than a vector store: context-window stuffing and vector retrieval both worked badly, but per-agent MEMORY.md and per-day diary files written at session end and read at session start worked well—plain markdown, no embeddings or recall scoring. Third, remote sandbox before human review: Atlas's early PRs passed local tests but broke in CI due to feature flags, third-party services, and real traffic, so monday gave each agent a remote sandbox per session where every PR auto-deploys, runs tests and checks, and replayed production traffic before requesting human review. Fourth, PR Guardrails—automated review against monday standards: every engineering standard (metrics tagging, feature-flag hygiene, Datadog usage, security boundaries, database and microservice conventions, test quality, documentation) became an automated reviewer wired to internal knowledge through monday's MCP servers. Roughly one in five PRs fails at least one standard and is bounced back, and human overrides run in the low single digits; tens of thousands of PRs are evaluated per month with hundreds of thousands of standard checks executed. Fifth, CoWORK—monday boards as the shared state layer: agents were opening PRs no team owned, so monday added the CoWORK workspace, placing agent tasks, status, blockers, and handoffs on the same monday board as the team's, making accountability a solved problem rather than a system problem. By late Q1 2026, the bottleneck shifted from code generation to human review. Morphex, monday's first fully autonomous engineering agent, works within the same repo, CI pipeline, Guardrails, and revert protocols as every Builder. Nineteen of every twenty Morphex PRs merge automatically by passing every gate, shipping to production with no human in the loop. Morphex opens more PRs per month than most engineers.

monday.com's deployment of AI agents reveals the gap between proof-of-concept and production scale. While early agent work was limited to L1 (AI as a pair programmer, with tools like Cursor) and L2 (reusable sub-agents in teams), the company has moved toward L3—fully agentic delivery—by solving operational problems that greenfield demos sidestep. The key insight is that agents work best when embedded in existing team structure and accountability systems rather than run as isolated background processes. By storing agent identity and state across Slack, GitHub, and monday boards, and by routing events through SNS and SQS to EKS-hosted agent pods, monday ensures that every agent has stable accountability and visibility to the humans they work alongside.

The five retrofits—evals before model upgrades, plain-markdown memory files instead of vector stores, remote sandboxes before human review, automated Guardrails checking, and CoWORK boards as shared state—each address a failure mode that emerged at scale. The adoption curve (nine in ten engineers using AI coding tools, up from roughly half a year ago) and throughput gain (per-engineer PR output up more than half) demonstrate that the system is not just theoretically sound but operationally delivering value. Morphex's 95% autonomous merge rate (19 of 20 PRs) suggests the next bottleneck may shift from human review to task selection and longer-horizon planning.

No comments yet. Be the first to share your thoughts!