Are Agents Just an Unreliable Intern?

Friends! Thank you for your patience as I work through some challenging ideas that I’m developing. Thomas Mann write that “a writer is someone for whom writing is more difficult than it is for other people.” And as we live in these unprecedented times, the act of curating ideas into writing that distills this moment and guides our thinking as we go forward is especially trying.

Building trust in AI systems has become a critical concern, with organizations implementing comprehensive governance frameworks, monitoring systems, and "onboarding" processes for AI agents similar to human employee training.

The thinking in these papers suggests that smaller, narrower AI agents and models with highly context-aware onboarding and training will improve reliability and trustworthiness and function as creative collaborators. This could be a valuable departure from the one-model-to-rule-them-all approach of the frontier firms.

These papers emphasize transparent, accountable AI systems with clear boundaries, escalation paths, and continuous feedback loops to ensure reliability and maintain human oversight.

Sources:

AI systems are increasingly designed to adapt and learn from individual users and organizational contexts, moving beyond one-size-fits-all solutions to highly personalized experiences. This includes AI that can understand organizational knowledge, adapt to specific workflows, and provide contextually relevant assistance while maintaining alignment with human values and goals.

Sources:

One failing of the ChatGPT-5 launch was a lack of product thinking; OpenAI misjudged their customers and how they use ChatGPT. Contrast that with Anthropic, who is taking a more customer-centric approach to developing their product lineup. The fact that they have a clear picture of their customer as the enterprise business operation is a big help. OpenAI’s grasp of their customer base is less clear.

As we explore this early relationship between humans and AI, the user experience is rapidly evolving from simple chat-based interactions to more sophisticated, context-aware systems that can understand intent, maintain conversation history, and operate across multiple modalities. This includes the development of AI systems that can better understand human psychology, emotional states, and cognitive processes to provide more effective assistance.

Sources:

Business leaders are making a big push to use AI, and that means the workplace transformation is less about eliminating job, but in reshaping job roles and required skills. The focus has shifted from job displacement fears to understanding how AI augments human capabilities and creates new types of work. Organizations are investing heavily in reskilling programs and developing new roles like "AI enablement managers" and "PromptOps specialists" to bridge the gap between human expertise and AI capabilities.

Sources:

There’s a growing stark gap between AI agent hype and reality. As workers are trying to delegate tasks to AI agents, 62% of these workers view them as unreliable. In fact, 54% say that agents create extra work rather than saving time. Top models achieve only 82% accuracy on document processing and below 70% on customer service tasks. Even the best solutions achieve below 55% goal completion rates with CRM systems, with only a 25% probability of completing all their test tasks. Key problems include shallow memory, inability to learn, compounding errors in multi-step workflows, and vulnerability to attacks. The industry is entering a disillusionment phase.

As we learn more about context engineering, as tools like Claude Skills enter the market, and as we learn more about how to manage scope, skilled users can improve agent performance, but agents are still in the early days, and agentic intelligence may simply not be ready for prime time.

Sources:

Justin Sung tested ChatGPT Study Mode, and his review revealed some important insights about how learning works with tools like AI. Here’s what we can learn about when — and when not — to use AI in the classroom, and what that means for how educators should maximize the benefit of classroom time.

Click here to read the full article.