João Freitas

The following is a comparison on how he would classify AI Agents in autonomy levels, just like self driving cars are classified.

Level 0 — Instruction-Driven Interaction

Traditional human-computer interaction. The system has predefined inputs and outputs. Program environments are static, and there are no self-learning or adaptive abilities. The majority of software today fits inside this bucket.

Level 1 — Assisted Cooperation

AI agents manage simple and predefined tasks, such as acting as active partners in task execution. AI-powered writing assistants like Grammarly fit in this category. They learn from user preferences but primarily rely on predefined rules and user confirmation.

Level 2 — Supervised Interaction

At this level, AI agents independently handle routine tasks within familiar contexts while humans supervise their actions. An example might be automated email filtering that sorts messages into different categories (spam, promotions, and important). They learn from user behavior but need human intervention.

Level 3 — Contextual Autonomy

Agents operate across diverse tasks within their defined scope and adapt based on experience. An example might be a customer service chatbot that can handle a variety of queries but needs human intervention for new or complex inquiries. These AI agents utilize tools such as external APIs or databases to complete their tasks. At this level, humans are primarily auditors of decisions made by the AI.

Level 4 — Monitored Interaction

AI agents at this level exhibit advanced problem-solving and learning capabilities, improving over time. They learn from interactions and refine their functionality but require occasional human intervention. These AI agents can utilize different tools and consistently break down problems into subproblems. Moreover, they can not only use existing tools but learn and develop new primitives.

Level 5 — Autonomous Intelligence: Governed Interaction

AI agents at this level operate with complete independence within their defined scope. They improve and innovate without human input. A (hypothetical example) might be an AI researcher that formulates original research questions, designs and executes experiments, interprets results, and publishes research papers without human intervention.

#reads #matt rickard #ai #autonomy