ITMC Innovationsforum
UniversitΓ€t Hamburg - Juni 2026

Building a Coding Factory

How AI Agents Are Reshaping Software Engineering

Marvin Kruse

Marvin Kruse

Technical Lead at OTTO

Hi, I'm Marvin

Marvin Kruse
Role Technical Lead at OTTO
Since June 2024
Experience >15 years β€” in various industries
Focus Platform & AI-Engineering, Developer Experience,
Architecture & Distributed Systems
Alma Mater Dipl. Wirtschaftsinformatik, Uni Hamburg πŸŽ“

Agenda

1

The Evolution of the Coding Setup

2

Internals of a Coding Factory

3

Demo

4

Key Takeaways & Outlook

Chapter 1

The Evolution of the Coding Setup

Three Eras of Coding

πŸ–₯️

Classic

Nerd + Coffee + Laptop + IDE

You write everything

πŸ€–

Modern

Nerd + Coffee + Laptop + IDE + Copilot

You write, AI assists

🏭

AI Coding Factory

Nerd + Coffee + Phone/Browser

You steer, AI writes

Your laptop is just the steering wheel β€” the engine runs elsewhere.

What Changed?

LLMs can write code β€” everyone knows that by now

But: Agents can also plan, structure, and execute code autonomously

The shift: from "AI as autocomplete" to "AI as a coworker who works independently"

How We Build Software β€” Today

πŸ’Ό Business Analyst requirement DEVELOPMENT TEAM πŸ‘€ PO writes ticket πŸ‘€ Dev writes code πŸ€– AI Copilot assists local dev loop (run, test, fix) Integration commit, push & open PR πŸ‘€ Dev reviews PR CI Pipeline Build, test, lint CD Pipeline Deploy to prod triggers πŸ‘€ Dev deploys pass fail

What if Coding Was a Factory?

πŸ‘€ Dev/PO/BA writes spec spec CODING FACTORY πŸ€– Agent writes code Agent N (parallel)... Integration commit, push & open PR CI Pipeline Build, test, lint fail closed loop πŸ‘€ Dev reviews PR pass request changes CD Pipeline Deploy to prod approve

Why Build Your Own?

Offload compute from the laptop

Agents work async β€” even while you sleep

Codex, Copilot Agent, Devin go there β€” DIY teaches you what's real

Self-experiment: how far can you push it today?

Chapter 2

Internals of a Coding Factory

The Factory β€” Schematic

CONTROL PLANE πŸ‘€ Me Phone / Laptop / Couch Happier Control & Monitoring Lateralus Orchestration (planned) GitHub Code, Issues, CI/CD spec EXECUTION PLANE Ubuntu VM Bubblewrap (Sandbox Container) Schism (CLI Shim) Agent (Kiro / OpenCode) Spec Engine Tools Feedback Loop starts agent git push + PR

The Control Plane

Happier β€” Start sessions, monitor progress, intervene (Web + iOS)

Lateralus β€” Orchestrator for multi-agent runs (planned)

GitHub β€” Source of truth: Issues, PRs, Code

You define what gets built. The control plane makes it happen.

Happier β€” The Control Interface

Web + iOS app β€” relay to the coding factory

Start sessions from a GitHub Issue or prompt

Monitor agent progress in real-time

Intervene: cancel, redirect, or approve

You'll see this in action in the demo.

The Execution Plane

Ubuntu VM β€” Homelab or cloud

Bubblewrap β€” Sandbox container, restricted network

Schism β€” CLI shim that blocks destructive/expensive commands (human-in-the-loop)

Agent (Kiro / OpenCode) β€” Reads spec, plans tasks, writes code, commits

Matroschka: VM β†’ Bubblewrap β†’ Schism β†’ Agent. Each layer adds safety.

Chapter 3

Demo

:(

Your PC ran into a problem and needs to restart.

We're just collecting some error info, and then we'll restart for you.

37% complete

Stop code: DEMO_GODS_HAVE_SPOKEN

Launch Demo β†’

Chapter 4

Key Takeaways & Outlook

β‘  We're Just Getting Started

You can build your own coding factory β€” today

Vendors are productizing it: Codex, Cursor, Github, Multica, etc.

The patterns are emerging β€” coding factories will become a commodity

β‘‘ The Skills That (Still) Matter

Systems thinking

Distributed systems & clean architecture

Unix fundamentals

Can't review what you don't understand

15 years of building systems without AI β€” that foundation still matters

β‘’ LLMs Are Not Intelligent

Stochastic β€” plausible, but no understanding

Security holes

Hallucinated libraries

Inefficient approaches

The agent is only as good as the senior who reviews

10Γ— faster code means
10Γ— faster problems β€”
in socio-technical systems, speed alone isn't the answer.

Thank You!

Danke Meme