Every Agent Needs a Box — Aaron Levie, Box

Companies must adapt their workflows to effectively utilize AI agents, rather than expecting agents to adapt to existing practices. Early adopters will gain a significant advantage and see compounding returns. However, deploying these changes across most companies will take time.

The increasing number of AI agents necessitates a robust infrastructure to ensure their effectiveness, governance, and security. Security incidents are expected due to prompt injection and unauthorized data access. Data governance, access controls, and regulatory compliance are crucial for managing the data that agents can access and the workflows they participate in.

AI coding has seen rapid adoption due to favorable conditions such as broad code base access, text-based medium, and strong model training. However, other areas of knowledge work face headwinds including limited data access, non-text mediums, and complex access controls. The rest of the economy will need to update its workflows to make agents effective, which will take multiple years.

Context engineering is crucial due to the limitations of context windows. Managing the vast amount of enterprise data with limited token capacity is a significant challenge. Models need to improve in their ability to determine when to stop searching for information and give up on a task when no answer can be found.

Unlike coding, knowledge work introduces new risks due to the "slop" factor, where AI models generate slightly different outputs each time. Constraining models to specific tasks and ensuring accuracy is essential. The lack of professional liability in engineering compared to fields like law and healthcare raises questions about accountability for AI agents.

Box partners with Apex to evaluate agents and uses its own internal evals with documents across various industries. These evals test the reasoning capabilities, compute capabilities, and context rot issues of different models. The company is seeing incredible jumps in model performance.

Box has a dedicated AI agent team supported by search and infrastructure teams. This team operates as an internal startup, driving innovation and addressing the existential need to get AI right. The broader company provides essential security, compliance, and governance features.

Box envisions agents both reading and writing data within its file system. While reading data presents challenges due to the vast amount of information, writing data is technically easier. The company aims to support any agent using Box as a file system for its work, providing a sandboxed workspace for collaboration and data storage.

The challenge is that information changes rapidly, requiring constant updates to AI systems. Companies need to capture and digitize tacit knowledge to shorten ramp cycles and reduce rework. A company culture of capturing and digitizing information in an agent-ready format is essential.

The founder balances delegation with direct involvement in critical areas, particularly those related to AI. The founder's experience and insights are crucial for navigating the complex landscape and making key decisions. The founder feels a personal responsibility for the company's success or failure in the AI space.

While process is important for managing routine operations, certain areas, like AI, require constant engagement and attention. The pressure to stay ahead in the rapidly evolving AI landscape creates a 24/7 environment.

The production function involves connecting the dots between internal work, external feedback, and industry trends. Building in public is a natural way to share insights, gather feedback, and incorporate new information.

The discussion explores the future of film and the potential impact of AI. While AI could democratize creativity and enable more people to make films, there are concerns about maintaining artistic quality and originality. The goal is to use AI to accelerate the production process and enable new forms of entertainment and art.

The discussion reflects on the influence of TechCrunch in the early days of Silicon Valley and explores how to create a similar launchpad for AI innovation. The key is to create an environment where creative people can connect and collaborate.

With the go direct model, every company needs to become a media company to communicate with its audience. DevRel is becoming increasingly important as companies need to find ways to get agents to see their stuff.

There is a huge demand for DevRel professionals, but a limited supply. The most talented individuals are often working for themselves in the creator economy. Companies need to take media seriously and invest in DevRel to attract and retain talent.