I Built a Self-Hosted AI Engineering Team That Won't Push Code Without My Approval

Description and use

A self-hosted AI engineering platform that sandboxed agents, enforces multi-stage reviews, and requires human approval before pushing code, ensuring security and quality.

Problem

AI coding agents often ship incorrect or insecure changes without human oversight, leading to broken code, accidental leaks, and unsafe dependencies.

Outcome

Implemented a secure, model-agnostic AI coding pipeline that prevents accidental pushes, catches secrets and unstable dependencies, and requires human approval before integration.

Workflow

• Define task and handoff notes
• Planner reads code and notes, writes plan with files, acceptance criteria, and risks
• Coder writes code in its own Git worktree
• Quick checks validate secrets, syntax, debug leftovers, and dependency stability
• If checks fail, coder revises code
• Reviewer reviews requirements, crashes, tests, and agreement up to 3 rounds
• UI and security critics run app, capture screenshots, and evaluate UI/security changes
• Git agent creates commit and PR
• Human approves via dashboard
• Branch is pushed to GitHub

Evidence

Agents run inside the coding harness's sandbox (Landlock/bwrap on Linux).

Source

The Planner reads the code, the project's handoff notes, and team memory, then writes a plan with files, acceptance criteria, and risks before any code is touched.

Source

Quick checks catch secrets, broken JSON/YAML/JS, debug leftovers, and risky new dependencies with plain code, no model call.

Source