I have seen the future of AI, and we are all managing agents.
We are telling them where to go. What to look at. We are answering their follow-up questions. We are correcting them when they make a mistake.
These were some of the talking points at last week’s AI Engineer conference in London, which brought together people from across the industry, including Google, Anthropic, and OpenAI.
The talks were varied, highly technical, and curated for the people steeped deepest in AI right now.
What was striking was how many of the presentations and discussions weren’t so much about the quality and abilities of AI models and agents — software that can perform tasks semi-autonomously — but the humans managing them.
Ryan Lopopolo, a member of OpenAI’s technical staff, called the moment early in the show. He said coding had changed dramatically in late 2025 because of advances in AI tools. The role of today’s software engineers, he said, is to steer and unblock agents.
That prompted several recurring questions: How much control should we cede to agents? What should agents look like? Should they delegate to other sub-agents? Is human language too limiting for telling an agent what you want it to do?
The event quickly started to feel like an MBA for the AGI-pilled. Everyone had a perspective on how agents should be managed. Words like “guardrails” and “context engineering” (a plan to have the agent perform optimally while burning fewer tokens) were everywhere.
These details matter because there’s an emerging consensus that 2026 will be an inflection point, when agents move from the experimental phase into one where they’re more reliable and leap from coding into other domains.
“I think 2025 was all about exploring, and 2026 is all about putting these agents into production,” said Anthropic’s David Soria Parra onstage. It’s not just coders who are going to have to think about these things: Parra said he expects we’ll soon see more “general agents that will do real knowledge worker stuff” such as financial analysis and marketing.
In this utopian work future, the agents are doing the grunt work for us — but they still need oversight. That means they need the right documentation, context, and guidelines to keep them from careening off course and doing things they shouldn’t.
It’s an irony of this moment that companies — including Meta, Google, and Amazon — are cutting management layers but may also end up turning everyone into AI supervisors. Individual contributors at tech companies, who once coded away without worrying about their direct reports, are now delegating and reviewing work done by AI.
Another big topic of debate was the amount of control we should give to agents — especially given how prone they are to breaking things. There was more than one dig at a recent disruption at Amazon caused by an AI coding assistant.
Mario Zechner, the creator of coding agent Pi, struck a more cautious message than most speakers. Agents learned on the internet, which is filled with a lot of garbage code, he said. He proposed a model for software engineers working with agents: use them sparingly, and don’t let them make decisions for you. “All of the decisions it makes are learned from the internet,” he said.
Agentcraft
Keeping tabs on agents means being able to see them, which prompted another interesting question: what should an agent look like?
One answer came via a showstopping moment from Monday.com’s Ido Salomon, who built a program called Agentcraft that, yes, displays functioning agents in a “Warcraft”-inspired environment.
The user can spawn new agents, prompt them as they would in any other AI interface, and there’s a handy way to cycle through agents with follow-up questions or that need your approval to execute a task. A heat map shows you if your agents are at risk of colliding — a problem that can occur when running multiple agents in parallel. This can happen if two agents are editing the same file at the same time, or both are tweaking different code functions that rely on each other.
Several attendees who spoke to Business Insider weren’t from major AI labs but companies big and small that are embracing agents in the workplace.
Yann Mainier, a senior engineer at Sky UK who attended the event, told Business Insider he wanted to learn more about how he and his team can build better agents.
“It’s more about: once you get agents, how do you make sure they are doing a good job? You can’t check them the same way you are doing with traditional software,” he said. For example, ask an agent to write the same function twice, and it might do it differently. “You need to have other ways,” Mainier said.
Managing agents may also require rearchitecting parts of the web to make it more legible.
Vercel CTO, Malte Ubl, said that in the week leading up to the AI Engineer conference, more than 60% of page views on Vercel.com were from agents.
“We have to consider another shift that the software itself is going to be used by agents now,” he said. When an employee proposes a new feature or interface, Ubl said he now asks a new question: “How does an agent use this?”
Have something to share? Contact this reporter via email at hlangley@businessinsider.com or Signal at 628-228-1836.
Read the full article here
