Box CEO Aaron Levie tells BI the fix for AI's 'context rot' isn't one super-agent

Aaron Levie, the CEO of Box, recently shared insights on the limitations of AI agents during an interview. He highlighted a significant issue known as "context rot," which occurs when these agents are inundated with excessive information. According to Levie, providing too much data can lead to confusion, resulting in suboptimal outcomes. Levie explained that as AI models process more information, they can become disoriented and may lose focus on the task at hand. This phenomenon is particularly evident when agents are overwhelmed by the volume of data in their "context window," which is essential for synthesizing information before generating responses. He emphasized that relying on a single super-agent is not the solution; instead, it's better to deploy multiple specialized sub-agents tailored for specific tasks. "You need to break apart the agents and the contexts they handle," Levie advised. He believes that having a fleet of agents, each with defined goals and relevant contexts, is the way forward. This approach contrasts with the overarching Silicon Valley aspiration for a singular AGI entity. Levie, who co-founded Box in 2005, predicts that the sub-agent model will shape the future of large-scale AI systems. He stressed the importance of accuracy in the data fed to these models, stating that precise information is crucial for optimal performance. However, he cautioned that too much context could hinder rather than help the AI's effectiveness. The tech industry is abuzz with advancements in AI agents, as companies race to implement them for complex, multi-step processes. For instance, Regie AI has developed "auto-pilot sales agents" to engage with potential buyers, while Cognition AI's Devin tackles intricate engineering challenges. Additionally, PwC has launched an "agent OS" to enhance collaboration among various agents. Despite the potential of AI agents to streamline workflows and learn from experiences, challenges persist. Experts have noted that as the number of steps increases, so does the likelihood of errors. A recent analysis from Patronus AI revealed that even a 1% error rate at each step could escalate to a staggering 63% chance of failure by the 100th step. However, implementing safeguards such as filters and rules can significantly reduce the likelihood of errors, underscoring the importance of continuous improvement in AI systems.