I’ve seen the future of AI, and we’re all managing agents.
We are telling them where to go. What should I see? We are answering their follow-up questions. Any mistakes have been corrected.
These were some of the topics discussed at the AI Engineers Conference in London last week, which brought together stakeholders from across the industry, including Google, Anthropic, and OpenAI.
The talks were wide-ranging, highly specialized, and carefully selected for those currently most deeply immersed in AI.
What struck me was that many of the presentations and discussions were not about the quality or capabilities of AI models or agents (software that can perform tasks semi-autonomously), but about the humans who manage them.
Ryan Lopopolo, a member of OpenAI’s technical staff, evoked this moment early in the show. He said that due to advances in AI tools, coding will change dramatically in the second half of 2025. The role of software engineers today is to guide agents and unblock them, he said.
As a result, several questions arose repeatedly. “How much control should I give to the agent?” What should the agent look like? Should it delegate to other subagents? Is human language too limited to tell the agent what you want it to do?
Anthropic’s David Soria Parra said agents are transitioning from coding to other jobs. hugh langley
The event quickly started to feel like an MBA for people studying AGI. Everyone had a point of view on how agents should be managed. Words like “guardrails” and “context engineering” (planning to make agents perform optimally while consuming fewer tokens) were thrown around.
These details are important because there is a growing consensus that 2026 will be the tipping point when agents move from an experimental stage to a more reliable stage and leapfrog from coding to other areas.
“I think 2025 is all about exploration and 2026 is all about getting these agents into production,” Anthropic’s David Soria Parra said on stage. Programmers aren’t the only ones who have to think about these things. Parra said there will soon be more “full-service agents doing true knowledge worker work” such as financial analysis and marketing.
In this utopian future of work, agents do the heavy lifting for us, but they still require oversight. This means you need proper documentation, context, and guidelines to ensure you don’t go off course or do things you shouldn’t.
The irony of this moment is that companies — Including Meta, Google, Amazon — is reducing layers of management, but they may all eventually become AI supervisors. Individual contributors at technology companies, who once coded without worrying about their direct reports, now delegate and review work done by AI.
Another big topic of discussion was how much control should be given to agents, especially given their tendency to break things. In Amazon’s recent disruption caused by AI coding assistants, there were multiple discoveries.
Mario Zechner, creator of the coding agent Pi, struck a more measured message than most speakers. He said the agents learned from the Internet, which is full of garbage code. He proposed a model for software engineers working in collaboration with agents. It’s about using agents sparingly and not letting them make decisions. “All decisions are learned from the internet,” he said.
agent craft
Monitoring an agent means being able to see the agent. So another interesting question arose: what should an agent look like?
One answer comes courtesy of a surprising moment from Monday.com’s Ido Salomon. He built a program called Agentcraft, which displays agents functioning in environments inspired by, well, “Warcraft.”
Users can spawn and prompt new agents like any other AI interface. It also provides a convenient way to cycle through agents who need approval to ask follow-up questions or perform tasks. The heat map shows whether agents are at risk of collision. Conflicts can occur when running multiple agents in parallel. This can occur if two agents are editing the same file at the same time, or if they are both coordinating different code functions that depend on each other.
How can you make controlling an agent fun? Make it a video game. hugh langley
Several of the participants Business Insider spoke to did not come from major AI labs, but rather from companies large and small that host agents in their workplaces.
Jan Mainia, a senior engineer at Sky UK who attended the event, told Business Insider he wanted to learn more about how he and his team are building better agents.
“The key is, once you hire agents, how do you make sure they’re doing a good job? You can’t check them the same way you can with traditional software,” he says. For example, if you ask an agent to write the same function twice, it might execute it in different ways. “There has to be another way,” Mainier said.
When managing agents, you may also need to redesign parts of the web to make it easier to read.
Malte Ubl, Vercel’s CTO, said that in the week leading up to the AI Engineer conference, more than 60% of Vercel.com’s page views were from agents.
“We need to consider another change where the software itself becomes used by agencies,” he said. When employees suggest new features or interfaces, Ubl says, they start asking new questions: “How will agents use this?”
Have something to share? To contact this reporter via email, please specify the following address: hlangley@businessinsider.com Or text us at 628-228-1836.