AI Camp Berlin · April 2026
We can each build anything.
So why can't we
build together?
🪜 The 5 Levels of AI Integration
L1 — Autocomplete
AI finishes your sentences. You still write everything.
L2 — Chat
Describe intent, AI generates. No execution, no verification.
L3 — Agentic
AI navigates, reads, plans, executes multi-step. You review.
L4 — Harness
AI writes, tests, reads failures, patches its own errors — on its own. You define inputs, outputs, and acceptance criteria. The bottleneck shifts from speed to spec quality.
L5 — Dark Factory
AI owns generation, testing, and deployment. No human in the critical path. Your job: write perfect specs and maintain the verification systems that catch what goes wrong.
NB
Framework by Nate B. Jones — The Nate Newsletter · Substack
✉️ L2 — Invoice automation
Drop a PDF. Never touch it again.
Vendor, date, amount, category — extracted and filed automatically. Zero infrastructure, zero cost.
Drive Inbox
drop PDF
→
Gemini Flash
extract JSON
→
Logic layer
route + dedupe
→
Sheets
vendor · amount · category
→
Archive
filed by year
Architecture
Google Apps Script + Gemini
Monthly cost
€0
Self-healing
learns new vendors
Audit trail
every decision logged
Spreadsheet output
MD5 dedupe · script_log audit · Drive verification
📈 L3 — Campaign analysis pipeline
😰 The problem
"The reviews that happened were symptom-driven — something had to go visibly wrong before it got attention."
- Everything else got a quick skim or nothing
- Advertisers got inconsistent advice
- Patterns across campaigns were invisible — nobody was looking across campaigns
Manual review per campaign: 40–60 min. Book of 200+ campaigns. The math doesn't work.
⚙️ The pipeline
"A pipeline that runs every night whether anyone asks it to or not."
- Every campaign, every night — fully headless
- Deterministic detection → Claude synthesis → recommendations
- 15-test E2E suite gates every deploy
"The quality floor went up because nothing gets skipped — not because the ceiling got higher."
📊
Campaign Analysis Dashboard
Every campaign, every night. Deterministic detection → Claude synthesis → recommendations. 15-test E2E suite gates every deploy.
100%
campaigns covered
15
E2E tests
0
manual steps
⚙️
L3 — The pipeline, live
🪤 Still L3 — the solo builder trap
I shipped it. It works. But I'm still the only reviewer, the only debugger, the only one who understands the system.
🧠
Context is unshared
The system lives in my head and 50 agent chat sessions. There's no onboarding path. Nobody else can pick this up.
🔍
Review is broken
It's not really "code" anymore. How do you review a diff when the agent wrote 80% of it and the logic is in the prompts?
⚡
Velocity doesn't transfer
I can ship a full feature on a Sunday. The moment a second person joins, we're back to meetings, alignment, "can you explain what this does?"
At L3 the AI is a teammate — but I'm still the single point of failure. That's the plateau.
⚙️ L4 — what it means
You stop managing the code. You start managing the system.
🤖
Self-running loop
AI writes, tests, reads failures, patches errors — on its own. No human in the middle.
📐
You own the spec
You define inputs, outputs, acceptance criteria. The system figures out how to get there.
📈
System compounds
Memory layer, self-optimisation, contradictions resolved automatically. Gets smarter without hand-tuning.
Ship while you sleep. Your job shifts to strategy, not execution. The team works on why and what.
At L4, the bottleneck is no longer you. It's the quality of your specs — and how well your team collaborates around them.
🪞 The L4 mirage
Consulting for FaustAI — an AI documentation startup doing impressive things. This looked like L4 from the outside.
What impressed me
- Agents running constantly in production
- Memory layer, contradiction detection, structured output
- Backend self-optimising prompts at scale
- Product defining inputs and expected outputs
Engineer defines the harness. Product defines the spec. Agents do the work. Looks like L4. ✅
The collaboration gap
- Only one person fully understands the system
- No closed feedback loop — output quality still requires hand-tuning
- Remove the engineer → everything stops
- Product can define what good looks like, but can't verify how to get there
The engineer is the harness. Smart system — but not yet a collaborative one.
🫀 Be more human, among bots
Some things I've tried. They help — but they don't fully solve it.
🎫
Open-ended tickets
Jira-skill staying human among bots: avoid overspecification, define problem + desired outcome, but leave more for teammates to fill in — AI can fill in the rest.
✓ staying human among bots
🤝
Backlog refinement
Open-ended tickets provide fertile ground for active discussion. Teammates — aka "agent-representatives" — discuss how the puzzle pieces fit together.
✓ slows the right things down
❓
Still unsolved
What does code review mean at L4? How do you onboard? Who owns a system that is mostly prompts?
⚠ no good answer yet
These are patches, not a system. The real question: what does team collaboration look like when everyone has an agent running?
🤔 Questions I'm carrying
What does an AI-native team actually look like?
01
How do you onboard someone into a project built by one person + 50 agent sessions?
02
What does "code review" mean when the agent wrote 80% of it?
03
Who owns the system when it's mostly prompts and harnesses?
04
Is a PM's job now: write better specs, run better refinements, stay out of the way?
We've solved individual velocity. We haven't solved collective intelligence.
🎤 Let's talk
Open facilitation — 15–20 min
🪜
Where are you on the ladder — individually vs. as a team?
🧩
Has anyone cracked the onboarding/context problem in an AI-native codebase?
🐢
What's the hardest thing to do together that used to be hard alone?