The gap between enthusiasm and execution was the focus of his keynote. Bousetouane began by asking the audience how many believed agents would transform their industries and change the way they work. Almost every hand in the room went up.
Yet despite that optimism, he argued that significant obstacles remain before AI agents can be deployed at scale. “The hype is there,” he said. “Everybody’s talking about Gen AI, LLMs and agentic systems, but with the hype, lots of confusion is emerging.”
Few agents reach production
According to Bousetouane, 79% of AI leaders say they have already embedded, developed or deployed AI agents. However, only two to four percent of AI agents and AI systems make it into production.
Part of the problem, he argued, is that organisations increasingly start with a technological solution rather than a business challenge. “You will get some use cases where your partner will start with the solution,” he said. “Use an LLM to solve this problem. Use an agent to automate this process.’”
There is no magic box or set of rules on how to screen, how to evaluate an agent.
Historically, businesses started by understanding the problem, the available data and the constraints before selecting a technological approach. The current rush to adopt AI is reversing that logic, placing pressure on organisations and engineering teams alike.
At the same time, Bousetouane noted that there are still no clear standards for building, orchestrating and deploying AI agents in a responsible and safe way.
The risks of autonomous systems
Among the technical challenges facing AI agents is what Bousetouane described as a “memory gap”. While humans interpret past experiences through cognitive signals, agents process previous interactions as text, making it more difficult for them to retain context.
Hallucinations remain another concern. While large language models can generate inaccurate information, Bousetouane warned that the consequences become more serious when autonomous agents are involved. “An agent impacts the environment,” he said. “Hallucination of an agent is problematic.” Unlike a chatbot, which can simply be prompted again, an agent can take actions and influence processes, systems and outcomes.
The challenge is compounded by a lack of accepted evaluation frameworks. “There is no magic box or set of rules on how to screen, how to evaluate an agent,” he said, pointing to hallucination, task success and drift as key areas to assess.
Building the infrastructure layer
Rather than focusing solely on individual agents, Bousetouane argued that organisations should think in terms of infrastructure. He described what he called a “harness” – a structured environment incorporating governance, guardrails, monitoring, observability and intelligence layers designed to support large-scale deployment.
Bousetouane also argued that agents may need to rely on different large language models for different functions, with one model used for reasoning and another for tool-calling. Not all models, he warned, are trained to understand a company’s private tools or proprietary data, which is where agents can make mistakes.
Without that infrastructure, agents often perform well during demonstrations but struggle when confronted with real-world complexity. “They are good in the labs, but when they go in the wild, they get confused,” Bousetouane said.
The future is agents. Agents are going to be the architects of our future.
Europe’s regulatory challenge
Technology alone will not solve the problem. Drawing on his experience working with more than 300 organisations, Bousetouane argued that product leaders, engineers and business teams frequently work in isolation from one another, creating challenges when trying to build scalable AI products. “Everybody’s building demos,” he said. “It works. It’s beautiful. But when we scale, it’s expensive. And it’s not safe.”
For European organisations, the challenge is heightened by regulation. Compliance with the EU AI Act, GDPR and other requirements must be considered from the outset. “We have to think of safety by design, security by design,” he said.
To address these challenges, Bousetouane presented the AI Steering Wheel, a framework developed with an AWS specialist in agentic systems. The framework is designed to support the full development cycle of generative AI products and agentic systems by aligning strategy, operations and engineering while clarifying ownership across teams. A book detailing the framework, use cases and implementation blueprints is expected to be published in the coming weeks.
A future web of agents
Bousetouane also introduced Proof Agent Harness, an open-source infrastructure for evaluating AI agents. The platform assesses agents across dimensions including task success, hallucination rates, policy compliance and regulatory requirements. Released just a week ago, the platform has already attracted more than 5,000 downloads, according to its creator.
Looking further ahead, Bousetouane predicted a future in which agents increasingly interact with one another across organisations rather than solely with humans. Such an ecosystem would require common protocols and trust mechanisms to enable autonomous systems to collaborate securely. “The future is agents,” he said. “Agents are going to be the architects of our future.”
In a final glimpse of that future, he suggested that agents may eventually possess their own identities, passports or digital credentials as they operate across a wider network of autonomous systems.
