AI Teammates: How monday.com Runs Production AI Agents on Amazon Bedrock¶

Ch04.680 AI Teammates: How monday.com Runs Production AI Agents on Amazon Bedrock¶

📊 Level ⭐⭐ | 4.6KB | entities/ai-teammates-mondaycom-production-ai-agents-bedrock.md

AI Teammates: How monday.com Runs Production AI Agents on Amazon Bedrock¶

Overview¶

AWS blog post (2026-07-22) by Claudio Mazzoni, Ofek Dayan, Netanel Abergel, Moran Zilberstein, and Erez Drutin sharing monday.com's production architecture for running AI agents on Amazon Bedrock at scale. Nine in ten Builders use AI coding tools monthly; per-engineer PR throughput is up by more than half.

Three Levels of AI Engineering¶

L1 (Assistant) — Engineers use AI as pair programmers: Cursor for fast reflexive work, Claude Code for heavy lifts; adoption nearly doubled YoY
L2 (Skills & Sub-agents) — Teams build reusable agents for repeated work; engineers remain in the driver's seat; per-developer PR throughput stepped up by more than half
L3 (Multi-agent) — Fully agentic; agents own delivery end-to-end while engineers orchestrate; agents take tasks from boards, communicate in Slack and monday, ship code alongside humans

Architecture: Sphera Agent System¶

The system uses seven AWS services: SNS, SQS, EKS, RDS, ElastiCache, EFS, and S3. AWS Secrets Manager handles per-session secrets; Amazon Bedrock handles model calls; the monday-agent-sdk runs inside each agent runner pod.

Event path: External triggers → SNS → per-team SQS queues → monday Builders CoWORK (SQS consumers on EKS) → resolves agent ownership → routes to agent runner pod. Pub/sub gives retries/DLQs, back-pressure, durable replay, and concurrent fan-out.

monday-agent-sdk: A thin wrapper around Claude Agent SDK providing provider neutrality (routes to Bedrock), cold-start cost optimization (warm node_modules), and a custom harness for evaluation, plugin composition, Slack/monday/GitHub integration, and output review standards.

State, Memory & Sessions¶

Three kinds of state with distinct stores:

Live state (ElastiCache): current task, run cursor, distributed lock, heartbeat, agent/human message log — sub-millisecond reads, self-expiring keys
Sessions and memory (EFS): each session is a directory (/sessions/<id>/ with repos/, secrets.json, messages/; /agents/<id>/ with MEMORY.md and diary/YYYY-MM-DD.md)
Durable records (S3): final transcripts, snapshots, artifacts, evals keyed by session ID

Five Production Retrofits¶

Evals before model upgrades — deterministic metrics (PRs merged, revert rate) + LLM-scored evals across 5 dimensions
Memory is a file, not a vector store — per-agent MEMORY.md + diary/YYYY-MM-DD.md; plain markdown, no embeddings
Remote sandbox before human review — PR auto-deploys to sandbox; tests and replayed production traffic run before review
PR Guardrails — every monday engineering standard automated as a reviewer; tens of thousands of PRs evaluated per month; ~1 in 5 PRs fails a standard
Builders CoWORK — monday boards as the shared state layer; agent tasks, status, blockers live on same monday board as the team's

Morphex: Fully Autonomous Engineering Agent¶

Morphex is monday's first fully autonomous engineering agent. 19 of every 20 Morphex PRs merge automatically using a confidence score combining: deterministic Guardrails outcome, per-agent eval trajectory, per-(agent × repo × change-class) historical revert rate, and sandbox outcome. Approximately 3 in 10 PRs merged, ~3/4 with zero human edits, revert rate in low single digits. Guardrails stopped ~1/4 of agent PRs before reaching a human.

Source¶

AWS Machine Learning Blog

关联¶

→ 原文存档 - 相关概念: Harness Engineering