We can each build anything. So why can't we build together?

Waldo Vanderhaeghen
#ai#engineering#teams#agents#reflection

Every room I walked into at AI Bar Camp Berlin was stuck on the same problem. Different stacks, different industries, different session titles — one question underneath: we can each build anything now, so why can’t we build together?

That was the title of my talk. I brought it as a question because I couldn’t answer it. Nobody else could either.

The framework I opened with

The framing I used: a five-level capability ladder from Nate B. Jones, from ‘AI completes tasks’ (L1) through ‘AI agents run autonomously’ (L5). The interesting transition isn’t L4 to L5. It’s L3 to L4 — from tool to system — and it’s not a tooling problem. It’s a systems problem.

My own journey

I shared where I’ve been on this ladder:

L2: Invoice automation — drop a PDF in a folder, it lands in the right row in a tax spreadsheet. This worked cleanly. Narrow problem, clear output, easy to verify.

L3: A campaign analysis pipeline — 181 ad campaigns, run headless every night, with automated QA gates. I’m squarely here. Claude Code is my interface. Most of my personal projects live in this zone.

L3 → L4: I’m figuring this out. The tooling is available. The gap is specification quality and knowing when to trust the loop.

The solo builder trap

Here’s what I hit that I suspect many people in that room also hit.

Solo, I move fast. Very fast. But that velocity comes at a cost: the context lives in my head, spread across fifty agent sessions, and nowhere else.

When I try to bring a second person in, we don’t move twice as fast — we move backwards for a week while I try to transfer context that was never written down because the agent and I figured it out together.

Three things are broken that used to just be inconvenient:

  • Context is unshared. Nobody else knows why the code does what it does. The agent knew, but it’s gone.
  • Review is broken. How do you review a PR where the agent wrote 80% of it? You’re reviewing the output of a conversation you weren’t in.
  • Velocity doesn’t transfer. The second person joins and you’re back to meetings, because the speed was personal, not structural.

The L4 mirage

Consulting with the Faust AI team while preparing this talk, I saw something building at L4 — agents in production, a memory layer, self-optimising prompts. Genuinely impressive. But when I looked closely, there was a pattern that concerned me.

Only one engineer understood the full system. There was no closed feedback loop with humans outside that engineer. Remove that person and everything stops — because the engineer IS the harness.

That’s not L4. That’s a more fragile version of L3, with better aesthetics.

Real L4 needs to be legible to someone other than its author. That’s the part nobody has solved.

The same problem, in every room

My session was in the morning. I spent the rest of the day walking into other rooms — and that challenge kept surfacing. Different framing, different industry, different stack. Every session circled it.

Strip the framings away and every session was circling the same two questions: where does shared truth live, and who keeps it true?

Jan-Peter Kalus from Helion ran a session titled “I’m Not a Developer with an AI Tool — I’m a PM with an AI Developer”. His point: the PM’s job is now to be the structured, reproducible interface between the human intent and the AI building it. The role hasn’t disappeared — it’s shifted entirely toward specification and ownership.

Markus Andrezak from überproduct put it even more directly: lots of AI is still about individual contributions — “my newly developed skill”, “my chief of staff”. But how does sharing of skills and context actually work across departments? He didn’t have a clean answer either.

Jakub Mierzewski from Atlassian ran a lightning talk on exactly the question I’d opened with: agents and people teams — how do you collaborate, and what are the top challenges? Different audience, that question again.

Tim Schikora from Aklamio went after the infrastructure layer: how do you handle context, skills, and project files across an organisation? Git isn’t accessible enough. GDrive isn’t structured enough for review. Does the answer require a custom Markdown editor that sits between the two?

And Dirk Bartels from Bain asked the strategic version: Agentic AI in SDLC — should we build more, or should we build better? AI tooling adoption is widespread. Results are lagging. The question is whether we’re reaching for the wrong lever.

It’s all about shared context, shared knowledge, shared data.

Some people think the answer is git-based — every piece of context lives in version-controlled files that agents and humans read from the same place. Others believe in vertical context managers — humans (often PMs, actually) who own the canonical understanding of a domain and promote it as best practice.

One framing that kept coming up: product managers are becoming context managers. The job isn’t just defining what to build — it’s maintaining the shared truth that agents can act on. The question is whether that truth is a single evolving document (Karpathy-style), or a constellation of files in disparate places that together constitute reality.

Nobody had a clean answer.

What did feel settled across the rooms:

  • Claude is the dominant tool (the echo was loud and unprompted)
  • We’re re-inventing SDLC patterns for agents — kanban boards, tech lead roles, design reviews — because the old patterns still make sense, even if the actors have changed
  • Prompt-engineering-based team setups feel like 2024 thinking; the direction is toward agents operating within proper engineering systems, not instructions masquerading as architecture

What’s still unsolved

I asked the room directly: has anyone cracked the onboarding problem? The context-sharing problem? Code review at L4?

Nobody has. A few had partial approaches. Most had workarounds.

The questions I’m still sitting with:

  • How do you review code when the agent wrote most of it and the conversation that generated it is gone?
  • How do you onboard a new engineer into an AI-native codebase where half the architectural decisions exist only in agent context?
  • How do you maintain ownership when agents can touch anything?

If you’ve got something working, I genuinely want to hear it.

The slides

If you want to go deeper on any of this, the live slides from my talk are here: waldo.vanderlore.de/talks/aicamp-2026/

The deck is interactive — it’s the actual version I presented.

One thing did get settled. The bottleneck isn’t code generation and it isn’t infrastructure — it’s the rate at which humans make good decisions. That was in every room, under every session title. Next year I’m coming back with one thing solved: a system where the decisions, not just the code, are legible to someone who isn’t me.