What I took from Web Summit 2026
Web Summit week brought two themes into focus, security around AI agents and data that's ready for agents to use. The best conversation happened outside.
I’m at Web Summit in Lisbon this week. Most of the first day’s programme was about AI, as expected. So far I’ve come away with two themes I expect to spend more time on with clients, and one conversation I’ll remember.
Security around agents
This is where I expect to spend more of my time. The industry is moving so fast that holes are opening all over the place, and the advice I give clients has to keep pace with it.
The first problem is identity. An agent running on a developer’s laptop usually runs as that developer. It has the same file access, the same cloud login, possibly even the same rights to the production database. Every command it runs is carried out with a person’s permissions.
It’s also recorded as that person. Giving each agent its own identity separates agent access from human access, so what an agent does can be traced and audited as the agent’s work. Nobody wants an agent’s mistake logged under their own name.
The second is secrets. API keys and passwords get saved to disk in a config file, sent to a model as part of a prompt, or committed to GitHub along with the code. None of this is new, but agents make it easier to do by accident.
Agents are useful enough that people will keep using them, me included. So the work is in the setup around them. Separate identities for agents, the least access that gets the job done, and a clear picture of what each agent reads, stores and sends, so secrets stay out of files and prompts.
Agents as consumers of data
The other theme was data. The next day’s programme was cancelled, so a few people got together and ran a webinar instead. One of the sessions was with Keewano, a company building a database designed for agents to query, and it prompted this section. Until recently, the consumers of a company’s data were people looking at screens and other software calling an API. Agents are now a third kind of consumer, and they behave differently.
A person asks one question and reads the answer. An agent working through a task might make dozens of requests in a few seconds, follow up on whatever it finds, and combine sources nobody expected to be combined. That raises two questions for any business with data worth querying.
One is performance. Can the infrastructure cope with that pattern of requests, and can it answer quickly and compactly? Everything an agent reads is measured in tokens, the units of text a model processes. Lean responses keep the token count under control, which keeps the agent fast and its running cost down.
The other is security. The data returned to an agent should be restricted to what the task needs. Personal information is the obvious example. If a task doesn’t call for it, it shouldn’t be in the response at all. The most reliable way to stop an agent doing something unforeseen with data is to never give it that data in the first place. Returning only what’s needed has long been good practice. With agents on the receiving end, it matters more than ever.
A conversation outside the venue
The part of the day I’ll remember best came after I’d left. While waiting for an Uber to the hotel, I got talking to an elderly local man, originally from Angola. We spoke in Portuguese.
He was worried about AI. What it’ll do to the economy, the lack of jobs, a poor future for the younger generation, and people losing their mental capacity.
I told him that I work in tech and share the same concerns, and that I think it’s healthy to have them. AI is an opportunity, but we have to be careful with it.
The last of his worries is one I think about regularly and I’ve deliberately started making sure I only outsource part of my day’s thinking to AI! Handing over all of it isn’t healthy, and I think businesses need to recognise that as they adopt these tools.
We talked for about ten minutes. By the end we were friends, and he was giving me cryptic pub quiz style questions. I was getting them right, too. In Portuguese, which isn’t easy!
References
- KeewanoDB overview, Keewano