SentinelOps

From issue to deploy — and incident to prevention — GitMind runs a full AI agent pipeline inside GitLab: plan, code, test, secure, document, and heal itself.

Comment

Inspiration

Every engineer knows midnight. That moment when production breaks, the on-call phone rings, and the same exhausted team that shipped the bug now has to debug it, document it, postmortem it, and prevent it — all while the next sprint is already piling up.

We asked: what if GitLab could heal itself?

GitLab already has everything — the code, the history, the pipelines, the logs, the issues. It just needed a brain. That's SentinelOps.

What it does

SentinelOps is a 12-agent autonomous AI pipeline built natively on GitLab Duo that runs your entire software delivery lifecycle — from issue to deployment — and then heals itself when production breaks.

Pipeline 1: Issue → Production (Dev Pipeline)

  • 🧠 Planner Agent — reads the issue, checks Risk DNA memory bank for dangerous files, breaks work into tasks, estimates risk score
  • 💻 Coder Agent — writes safe implementation code, avoids patterns that caused past incidents
  • 🧪 Test Agent — generates unit + integration tests, adds extra coverage for HIGH RISK files
  • 🔒 Security Agent — scans for vulnerabilities, blocks merge if critical issues found
  • 📚 Docs Agent — writes changelog, inline docs, README updates automatically
  • 👁️ Review Agent — summarizes everything for human approval, gives GO/NO-GO recommendation
  • 🚀 Deploy Agent — checks all agents ran, verifies Risk DNA, ships to production only if safe

Pipeline 2: Incident → Prevention (Ops Pipeline)

  • 🔍 Investigator Agent — digs through logs, commits, pipelines to find exact root cause
  • 📝 PostMortem Agent — writes full incident report automatically
  • 🔗 Memory Agent — links incident to similar past failures, finds patterns across incidents
  • 🛡️ Prevention Agent — creates actionable tickets to stop it happening again
  • 🌱 Carbon Agent — measures environmental cost of incident, tracks LLM energy usage

The flywheel: Every incident feeds learnings back into the Planner for the next sprint. The system gets smarter with every failure.

How we built it

Every agent is a custom GitLab Duo agent with:

  • A specialized system prompt defining its role and decision logic
  • A curated set of GitLab tools (get_issue, create_issue_note, get_repository_file, list_issues, find_files, etc.)
  • Access to the Risk DNA Memory Bank — a JSON knowledge base stored in the repo that tracks historically dangerous files, past incident patterns, and deployment blockers

Agents communicate by posting structured comments on GitLab issues and reading each other's outputs. No external infrastructure needed — it all runs inside GitLab.

The model powering every agent is Claude by Anthropic, accessed via the GitLab Duo Agent Platform.

Challenges we ran into

  • Building agent-to-agent communication purely through GitLab issue comments with no external message bus
  • Designing the Risk DNA memory bank so agents could read and write shared institutional knowledge across sprints
  • Making the Deploy Agent truly safe — it must verify that every upstream agent ran before giving GO
  • Keeping all 12 agents coordinated without a central orchestrator

Accomplishments we're proud of

  • 12 fully working specialized agents, all public in GitLab AI Catalog
  • A self-healing ops pipeline that activates automatically on incidents
  • Risk DNA memory that carries institutional knowledge across sprints
  • Carbon tracking — we measure the environmental cost of every incident
  • Zero external infrastructure — runs entirely inside GitLab

What we learned

GitLab is not just a code host — it's a complete operating system for software teams. With the Duo Agent Platform, you can encode your entire engineering culture into agents that enforce your best practices automatically, every sprint.

What's next for SentinelOps

  • Auto-trigger agents via GitLab webhooks (no manual @mention needed)
  • Multi-project Risk DNA sharing across teams
  • Carbon budget alerts that block deploys exceeding energy thresholds
  • SentinelOps as a GitLab CI/CD template anyone can install in one click

Built With

  • anthropic
  • anthropic-claude-api
  • claude
  • devops-automation
  • gitlab-api
  • gitlab-ci/cd
  • gitlab-duo-agent-platform
  • gitlab-duo-chat
  • gitlab-webhooks
  • json-(risk-dna-memory-bank)
  • multi-agent-system
  • python
  • yaml
Share this project:

Updates