Building AI agents: the hard truths

The tech industry is buzzing with possibilities around AI agents – autonomous systems that can perform tasks on our behalf. In a recent talk, Cloudflare's Rita Kozlov cuts through the hype to address the genuine challenges of building practical, effective AI agents. Her insights reveal that while we've made tremendous progress with large language models, creating truly useful autonomous agents requires solving complex problems beyond the capabilities of current systems.

Key Points

• AI agents need to handle context management across conversations and tasks, but current approaches like chaining LLM calls or "brain dumps" have serious limitations in preserving state.

• Building agents that can interact with the real world through APIs requires addressing authentication, rate limiting, and the lack of a standardized way for LLMs to discover and use third-party services.

• Testing and evaluation of agents is uniquely challenging because success criteria are often subjective, and simulating realistic user behavior requires sophisticated approaches beyond traditional unit testing.

• Security considerations for AI agents are substantial, as they potentially have access to both sensitive data and control over critical systems when acting on a user's behalf.

When Agents Meet Reality

The most compelling insight from Kozlov's presentation is how building functional AI agents exposes the gap between theoretical capabilities and practical implementation. While LLMs can generate impressive outputs in controlled environments, creating autonomous systems that reliably perform useful tasks across different contexts reveals fundamental challenges in areas like context management, tool integration, and security.

This matters tremendously as businesses increasingly look to AI agents as the next frontier of automation. The industry has rapidly progressed from chat interfaces to agents that can take action, but many companies are discovering these hard truths only after significant investment. Understanding these challenges upfront helps organizations set realistic expectations and properly scope their agent projects.

Beyond the Talk: Real-World Implications

What Kozlov's presentation didn't fully explore is how these challenges are playing out in early agent deployments. Microsoft's Copilot initiative offers an instructive example – while heavily promoted, users report inconsistent results when Copilot attempts to execute multi-step tasks across applications. This illustrates precisely the context management and tool integration issues Kozlov highlighted. The agent frequently loses track of complex tasks or misinterprets the appropriate actions to take when jumping between applications.

Similarly, OpenAI's