Stacks/Internal engineering platform
monday.com Infra Agents Platform
monday.com built a shared platform where every infra agent is just a prompt, a tool allow-list, and a subscription, and the guardrails live in the platform: one-shot Kubernetes Jobs with per-run credentials, wrapped tools, an independent Bedrock risk classifier, trust-levelled Cognee memory, CI evals, and LangSmith traces. One quarter in, agents resolved about 50% of an average 1,688 Infra tickets a month without human escalation.
- Use case
- LLM agents that handle Infra tickets from R&D: reading and proposing changes to Terraform, Helm, GitOps, and CDKTF, reviewing infrastructure PRs, and resolving or escalating tickets
- Category
- Internal engineering platform
- Team size
- Built by the Production Engineering team; now 7 agents serving 6 Infra teams
- Orchestrator / framework
- An always-on TypeScript orchestrator receives monday.com webhooks and Slack mentions, matches them to a subscription, creates scoped credentials, and launches a fresh Kubernetes Job per run. A Python runtime inside the Job runs the tool loop (iteration cap 20 by default) through a policy dispatcher with wrapped tools and no generic shell or HTTP tool. Each agent is three config files: agent-type YAML, subscription YAML, and a Markdown system prompt.
- Memory
- Two Cognee datasets with trust levels: run observations (context only, never authoritative) and operator facts written only through a team-gated process. A scrubber removes secret patterns by regex and PII with Presidio before anything is saved.
- Evals
- Every PR touching a prompt, agent type, or the runtime runs an offline eval harness: fixed cases replayed through the real agent loop with fake tools, scored by assertions and an LLM judge. A regression fails CI. A separate Bedrock call with fresh context scores each run's risk from the raw diff and plan output.
- Observability
- Each run has one parent trace in LangSmith plus JSON logs to stdout; every boundary-crossing call is traced. Kill switches and a system-prompt override sit behind feature flags.
- Cost (stated in writeup)
- Not stated
- Cost per output (stated)
- Not stated in the writeup
- WhatWeRun estimate
WhatWeRun estimate
Not enough info to estimate
The post states volume (about 1,688 tickets a month, 20-iteration cap) but names only 'a Claude model on Bedrock', with no model version or tokens per run.
Reviewed Nov 13, 2026
- Observability tools
- Other tools
- CogneePresidioGitHub App tokens (per-run, ~1 hour)Slackmonday.com boards
- Source
- https://engineering.monday.com/how-we-gave-ai-agents-the-keys-to-production-and-still-sleep-at-night/
Pablo Sebastian Martinez · Published Oct 26, 2026 · Last verified Nov 13, 2026
Related
More like this
aws.amazon.comNatera SWARM lab-operations agent
Diagnosing lab instrument support tickets for Natera's cfDNA testing labs: a structured diagnostic report on every Jira ticket before an engineer opens it, plus interactive troubleshooting
blog.duolingo.comDuolingo's production agent platform
An internal platform so any Duolingo team can define a production agent once and invoke it from Slack, internal sites, a CLI, or another Temporal workflow. In use for fixing CI failures, addressing code-review comments, and a Slack bot for release managers that investigates crashes and summarizes findings