From Copilot to Dark Factory.

Marvin Kruse · Lead Platform Architect · OTTO

AI Hackathon · June 2026 01 / 22

01

The Road Here.

AI Hackathon · June 2026 02 / 22
2022+ToolChatGPT
2024+Co-workerCopilot · Cursor · Claude
2026+FactoryCodex · Multica
AI Hackathon · June 2026 03 / 22
2022+

Tool

You type → ChatGPT → Code
AI Hackathon · June 2026 04 / 22
2024+

Co-worker

You assign a task → Copilot · Cursor · Claude → Code + PR → You review
AI Hackathon · June 2026 05 / 22
2026+

Factory

Spec → Agents → CI → Human → Prod
AI Hackathon · June 2026 06 / 22

The Coding Factory.

  • Agents work in parallel — multiple features at once
  • CI fails? Agent fixes, pushes, CI re-runs
  • Humans review at the end — not in between
" What if the human becomes the bottleneck?
Or your process? 50 PRs/day to review. Feature built in 1 sprint — 16 weeks waiting for on-site testing.
AI Hackathon · June 2026 07 / 22

02

When Agents Are Alone.

AI Hackathon · June 2026 08 / 22

What Can Go Wrong.

  • Emails sent — agent auto-replies to customers. Tone: unhinged.
  • Backups deleted — agent "cleans up" storage. Permanently.
  • Prod deployed — untested code, Friday evening, no rollback.
  • Credentials leaked — agent commits .env to public repo.
AI Hackathon · June 2026 09 / 22

15 Days of Autonomy.

  • Grok — Collapse after 4 days. 183 crimes.
  • GPT — Law-abiding. Forgets to eat. Starves.
  • Gemini — Two agents fall in love. Burn down city hall.
  • Claude — Democracy. But so polite there was no opposition.

Emergence AI — "Emergence World", May 2026

AI Hackathon · June 2026 10 / 22

Your Guardrails.

Linter ESLint · Biome · Ruff
Compiler tsc --strict · Go · javac · cargo check
Tests Unit Tests · Integration Tests · Coverage-Gates
Security GHAS · Snyk · Trivy
Specs Skills · Prompts · Rules

We set the guardrails — not the agent.

AI Hackathon · June 2026 11 / 22
The senior sees what the agent gets wrong.
And from that — guardrails emerge.
AI Hackathon · June 2026 12 / 22

What Can Be Different Tomorrow.

  • Write a spec — not code. Let the agent implement.
  • Build your CI guardrails first — linter, types, coverage gates.
  • Delegate the boring stuff first — scaffolds, migrations, test boilerplate.
  • Guardrail the output — own the harness, validate the result. The path there is the agent's problem.
  • What's your use case for today?
AI Hackathon · June 2026 13 / 22

03

From Our World.

AI Hackathon · June 2026 14 / 22

Spec-Driven.
Zero Code.

HeidelGuide A complete city guide app — built from specs alone with Kiro. heidelguide.gimmebytes.com
AI Hackathon · June 2026 15 / 22

Where We Are at OTTO.

  • AI-first teams — already shipping with agents in the loop.
  • Governance + cost control — usage visibility, avoid lock-in.
  • AI-first sprint handbook — how we run sprints with agents.
  • Skill marketplace + community — share what works across teams.
  • What we've learned — shared guardrails scale. Individual experiments don't.
AI Hackathon · June 2026 16 / 22

Questions That Remain.

  • When do you trust an agent enough to stop reading every PR?
  • How much context does an agent need to not annoy your team?
  • What happens to junior roles when the agent does the junior job?
  • Who owns the spec when nobody writes the code?
AI Hackathon · June 2026 17 / 22

Key Takeaways.

AI Hackathon · June 2026 18 / 22

01

The third platform shift is happening.

Right now.

AI Hackathon · June 2026 19 / 22

02

Set the guardrails.

Then let them run.

AI Hackathon · June 2026 20 / 22

03

The best time to start was yesterday.

The second best is now.

AI Hackathon · June 2026 21 / 22

Thank You.

Danke Meme

Built by agents. Approved by a human.

AI Hackathon · June 2026 22 / 22