# Shipping software with AI agents, safely and verifiably

- **Client:** Mardi Gras Developers (our own delivery pipeline)
- **Sector:** Software development
- **Period:** 2026 – present
- **Repositories:** 7
- **Approval:** Always by a human
- **Preview:** One environment per pull request
- **Deployment:** Only via the pipeline, after merge

We built our own environment in which AI agents prepare software changes across seven repositories. An agent creates an isolated working copy, makes the change, runs the checks and offers the result as a pull request with a live preview. Merging and deploying only happen after a human approves.

## The challenge

- AI agents can do a lot of work, but an agent with too many permissions is a risk to code, data and production.
- Agents read web pages and files that may contain an attacker's instructions.
- Clients without a technical background wanted to request changes to their website themselves.

## What we built

- **One controlled route to production.** An agent works in its own working copy and cannot push, merge or deploy by itself. One command runs the checks, creates the commit and opens the pull request.
- **The right model per task.** The heaviest model does analysis and design, a second model does a critical review, and small mechanical changes go to a cheap model. Every result is checked against the repository's tests.
- **A preview per change.** Every pull request gets its own environment on its own web address, so you see the change working before it goes live.
- **Security in the environment itself.** Agents are not allowed to change production data and do not deploy themselves. Destructive actions require a human and a separate key. Web content and files count as untrusted input, and every mutating action is in an audit log.
- **Changing by conversation.** A client describes a website change in a chat. The environment turns it into a pull request with a preview and sends the link back. A human approves.

## Outcome

- Changes arrive as small, checked pull requests with a working preview.
- Agents cannot change production data or publish anything without human approval.
- Clients request changes in plain language and review the result in their browser.
- We set up the same arrangement for clients who want to use AI agents safely in their own development process.

## Technology

AI agents, MCP, Git worktrees, GitHub, Coolify, Cloudflare Pages, Docker, Playwright, Sentry, Linux

## Services involved

- [AI solutions and agents](https://mg-devs.nl/en/services/ai-solutions/)
- [AI infrastructure and security](https://mg-devs.nl/en/services/ai-infrastructure-security/)

---
Source: https://mg-devs.nl/en/cases/shipping-software-with-ai-agents/
