What I took home from AI Summit Barcelona 2026

I spent 22 and 23 September at AI Summit Barcelona. The logistics couldn’t have been worse if someone had planned it that way: queues for everything, and a sales pitch at the end of most of them.

I won’t go back, and I’d strongly advise you not to go either. But I’m glad I went, for reasons the conference itself had nothing to do with.

Two days in queues

I missed the opening keynote and the first panel discussion because I was still in line to check in. The organisers announced more than 10,000 attendees, and it felt like every one of them was in that line.

Inside, it didn’t get better. Anything worth seeing had a long queue before it, and getting to the front didn’t mean you’d see or hear anything. The rooms were packed far past capacity.

The workshops were built for about twenty people. The ones I tried to get into had more than a hundred attendees, most of them standing at the back with no way to follow along.

The few times I did manage to get a seat, most sessions were only fifteen minutes long. That’s barely enough time for a speaker to introduce themselves, let alone get into a topic. The speakers clearly knew their field deeply but the format gave them no chance to share any of it.

Sadly, the bulk of the talks weren’t talks at all. Most of the time, a sales representative would take the stage and walk through their product. I’ve nothing against vendors at a conference, but when the stage is for sale, there isn’t much left for the audience.

I’d paid extra for a gold ticket, which came with access to a networking area. This turned out to be three small tables squeezed into a corner between the vendor stalls… Zero networking happened there. Most of the time, the chairs were taken by groups of friends resting their legs.

On the bright side, the conference app was a nice touch. It gave a quick overview of the sessions and let you build your own agenda. However, it also sent a steady stream of push notifications advertising rental bikes.

My gold ticket cost €635, and the people running the event on the floor were volunteers. Plus, I had to pay extra for coffee inside the venue.

Put that next to a stage sold to vendors and aggressive ads in the agenda app, and a more cynical person would call the whole event a money squeeze. I’ll settle for saying that every decision I could see put the organiser’s revenue ahead of the people who’d paid to be there.

I’d give a first-year event more grace. This was its second year.

I want to be clear that none of this was the staff’s fault. They were all extremely helpful and kind throughout, while taking complaints from other attendees about decisions they’d had no part in. I talked to a few of them one evening, and from their side it had been chaos from start to finish.

The people made it

What saved the two days was everyone else.

I met someone who runs the .pe top-level domain, and we talked about how a national domain actually runs as an international operation. With a developer from the US, backpacking through Europe, I got into how organisations can be social constructs and still be completely real.

I met people from e-commerce and consultancy, from startups and from huge corporations, and we traded what’s working for each of us with AI; from the tools themselves to how to relate to it all.

That’s where the rest of this piece comes from. Most of what I brought home came from the people I got the opportunity to meet and talk with.

Nobody shares their context

That chat about organisations as social constructs applies here. An organisation exists only because people agree that it does, and it’s still real. What an organisation knows works the same way. It lives in people’s heads, and it’s real enough to decide whether a project succeeds.

Sharing that knowledge is where I saw the biggest gap.

Anyone who’s worked with agents for a while has learned two things: 1) curating and enriching an agent’s context makes it far more effective, and 2) too much context becomes a problem of its own.

So everyone builds their own answer: a local knowledge base, a retrieval process, notes on what the agent needs to know. Each person gets better at it, but the organisation gains nothing, because none of it is shared. Your collague has solved half your problem, and neither of your agents knows.

It’s an old problem. We’ve tried to document our work for as long as there’s been work to document, and we’ve never quite managed it.

I think it can be solved now, for one reason. Documentation always failed because the person writing it paid the cost while someone else got the benefit, later, if ever. Context written for an agent pays off straight away, for the person who wrote it. As a bonus, AI also makes this easier.

People are already doing this work, they’re just doing it in isolation.

The hard part is the second lesson. Sharing everything indiscriminately with everyone would bury our agents in useless context.

Kind of like how the call to “break down the silos” in organizations has been a mantra for decades, but it doesn’t work. The silos exist for a reason: the context that matters to one team is noise to another.

Most of what we know about building a product and running a business, we find out as we go, and it lands in different layers. Some of it holds for the whole organisation. Some belongs to a team, a company function or a single project. Some is one person’s preferences.

The tools I’ve found for this so far offer one big shared pool of context for everyone to draw from, with no sense of those layers. What’s missing is a way to share selectively: to know which parts of that knowledge belong in your agent’s context, for this task, right now.

Voice is where the vendors are heading

If the vendors on site are any sign, voice is the next interface for AI products. Stall after stall was selling some kind of voice agent.

It makes sense. The chat box was the first way we learned to interface with these systems, but typing is a narrow way to exchange information and instructions. Voice is a very plausible next step.

What interests me is why it might work so well.

AI is nowhere near replicating what it’s like to sit across a table from another person. But we already accept a degraded version of this all the time, as we talk to each other through a chat or a phone line. Inside that narrower channel, an agent has far less to imitate.

That’s the idea Turing’s test was built on. He had the conversation typed, so the machine only had to pass as a person in text. A phone call widens the channel, but not by much. An agent doesn’t need to pass as human in the room with you. It needs to pass on the phone, and that’s much easier.

Where I’m less sure is whether the vendors are right about the direction. A vendor floor at a conference only really shows who had the budget for a stall, which isn’t the same as where the field is heading. And the pitch I kept hearing most was “customer service that sounds human”.

In the EU, that agent now has to tell you it’s a machine. Since 2 August this year, the AI Act has required any system that talks to people to make that clear, unless it’s obvious anyway. So the call can be good enough to pass as a person, but the agent still has to say it isn’t one.

Review is the new bottleneck

I met a dozen people who’d hit the wall I keep hitting myself: agents now write code faster than anyone can review it.

Qodo’s 2026 report on AI code quality, published on the last day of the summit, puts a number on it. 26% of developers and 26% of engineering leaders named reviewing and validating AI-generated code as their top constraint, ahead of trust, workflow integration, security, and cost.

Writing the code used to be the slow part, and review was the check at the end. Now writing is close to free, and a person’s attention is the scarcest thing in the loop.

My own answer so far is a growing set of documented principles and heuristics for what makes code promising or troublesome. A reviewer agent applies them to each change, fanning out to as many as a dozen subagents that each cover one area: security, UX, data governance, performance, etc. Then I get back a summary of the change’s quality.

None of that takes the responsibility off of me. I still have to decide what good and bad code are, in rules clear enough for an agent to apply, and that’s painstaking work that I don’t see ever being done.

And I still have to review every change. The summary, however, plus a few random samples of my own, makes that a lot easier than reading all of it.

The fundamentals haven’t moved

Every developer I talked to was producing virtually all of their code with a coding agent, or several at once in most cases. Some of the setups people described were seriously impressive.

Many of them also felt some sadness about losing the craft. I can relate.

What got most of us into this was the problem solving — digging into a complicated system until an elegant solution turns up. That hasn’t gone anywhere, though. It just sits a level or two higher than it did.

What we’re losing is writing the code by hand, but the fundamental principles of product development and technical excellence were never in the typing.

AI changes how much one person can build, and how fast, by an order of magnitude. I produce in hours what used to take me days. That’s an enormous change in degree, but not in category.

You still have to find a real customer problem and solve it. You still have to communicate why your solution is worth paying for. The technical solution still has to fit the business problem. And you still have to break it into parts small enough to understand, joined through well-defined interfaces.

I’ve been developing software since 1996, and nothing in that time has changed work the way AI has. But what changed is who writes the code, and how quickly. What makes it the right code hasn’t changed at all.