The Ugly Truth About AI Agents (and How to Spot the Real Ones)

Many companies are 'agent washing' by falsely claiming their products are AI agents, but a true AI agent is autonomous, takes observable action, perceives its environment, adapts from feedback, and is proactive.

• Most AI agent claims are 'agent washing'; true agents are autonomous, active, perceptive, adaptive, and proactive.

• Verify an AI agent's capabilities against the five key characteristics to avoid being misled.

• Prioritize an AI agent's proven usefulness and reliability, as technical definition alone doesn't guarantee success.

The market is rife with 'agent washing,' where companies mislabel old features as AI agents, with only a small fraction of vendors offering genuine solutions. A true AI agent exhibits five key characteristics: autonomy, observable action, environmental perception, adaptability from feedback, and proactivity. While tools like ChatGPT Deep Research and Custom GPTs may possess some of these traits, they often lack others, highlighting the need for a clear understanding of agent capabilities to avoid being misled by marketing claims.

Defining a True AI Agent

• 00:01:33 An AI agent goes beyond a simple chatbot by perceiving its environment, reasoning about context, and taking concrete actions, such as sending an email rather than just drafting it. Major players like Amazon, Anthropic, Google, and Microsoft agree that true agents possess five characteristics: they are autonomous, take observable action, perceive their environment, adapt from feedback, and are proactive without prompting.

Identifying False Agents

• 00:03:48 Companies often engage in 'agent washing,' misleading customers about the capabilities of their products. For example, Presto Automation faced SEC charges for falsely claiming its AI voice tools eliminated human order-takers in drive-throughs. Such examples highlight how purported agents may fail the autonomy test, requiring more human intervention than initially promised and making it crucial to verify whether a product truly delivers on its 'agent' claims based on the five core characteristics.

Evaluating Specific AI Agents

• 00:04:43 QuickBooks Online's accounting agent successfully passes all five criteria, autonomously categorizing expenses, taking action by recording them, perceiving financial data, adapting to user corrections, and proactively performing tasks. Microsoft Copilot Studio agents also meet most criteria, being autonomous, perceptive, action-oriented, and proactive, though adaptability often requires manual building or human input. In contrast, Custom GPTs are autonomous, perceptive, and can take action but typically lack true adaptability over sessions and are not proactive.

The Importance of Agent Usefulness

• 00:08:12 Beyond meeting technical definitions, the ultimate utility of an AI agent is critical, as reliability is paramount for businesses. Agents, based on probabilistic technology, must consistently perform their tasks well. If a task does not require text or image interpretation/generation or a dynamic workflow, an alternate, simpler technology might be more appropriate than an AI agent.