What defines a true AI agent? A closer look at the spectrum of autonomy

Autonomous AI systems are being marketed heavily by tech leaders, with varying degrees of actual capability and autonomy.

Current landscape: The tech industry is experiencing a surge of interest in AI agents, with major figures like Jensen Huang, Mark Benioff, and Satya Nadella promoting their transformative potential.

• Nvidia’s CEO Jensen Huang describes AI agents as a “multi-trillion-dollar opportunity”
• Salesforce’s Mark Benioff views agents as “what AI was meant to be”
• Microsoft’s Satya Nadella suggests Software-as-a-Service (SaaS) is being superseded by agent-based approaches

Core capabilities: True AI agents must demonstrate specific design patterns and fundamental capabilities that set them apart from simpler automated systems.

• Planning, reflection, collaboration, and tool use are essential characteristics of genuine AI agents
• Agency refers to an AI system’s ability to control and direct its own program flow independently
• Autonomy represents the AI’s capacity to operate effectively across various contexts without human intervention

Market reality: Many current offerings labeled as “agents” fall short of true agency and autonomy.

• Numerous SaaS products market “agent-like” features that are actually limited automated workflows
• Many solutions are simply LLM prompts embedded in deterministic process flows
• These “agent-ish” systems often lack the broad contextual understanding and decision-making capabilities of true AI agents

Autonomy spectrum: A clear framework exists for categorizing AI systems based on their level of autonomy and agency.

• Level 0: Fully manual human operations
• Level 1: Basic rules-based automation
• Level 2: Systems using machine learning for specific tasks
• Level 3: AI operators with limited agency within defined parameters
• Level 4: True AI agents with broad contextual understanding
• Level 5: Theoretical artificial general intelligence (AGI)

Real-world applications: Genuine AI agents are emerging in specific domains while “agent-ish” solutions serve intermediate needs.

• True AI agents like Devin for programming and AI Scientist for research demonstrate advanced capabilities
• Enterprise applications include drug discovery, customer verification, and complex data analysis
• Many organizations currently benefit from hybrid solutions combining Level 2 and Level 3 autonomy

Future implications: The evolution of AI systems suggests a hierarchical structure where different levels of autonomy will coexist and complement each other, though widespread adoption of true AI agents remains a future prospect.

• Organizations must carefully evaluate vendor claims about AI agent capabilities
• “Agent-ish” systems serve as stepping stones toward more advanced autonomous solutions
• The path to full AI agency will likely be gradual and require careful integration of various autonomy levels