Introducing the HIA Agent Scale: A Better Way to Describe AI Agents
Everyone seems to be talking about AI agents. The trouble is that the term now describes everything from a chatbot that answers questions to a system that can plan and execute work with very little human involvement.
Those are very different things, especially if you’re deciding what an AI system should be allowed to do.
At Hybrid Intelligence Advisors, we wanted a consistent way to describe these differences. So we reviewed the existing frameworks and classification systems to see whether one could serve that purpose.
Several stood out. The MIT Media Lab’s NANDA framework offers a useful progression of agent autonomy, with particular attention to human supervision. Anthropic has done important work on agent behavior, autonomy, and safeguards. Microsoft’s agent maturity framework provides a broader view of organizational readiness and governance.
Each contributes something valuable. But we kept running into a gap.
Most frameworks emphasize how independently an AI system operates. They ignore the role it plays in its relationship with a person.
An AI that answers questions and one that challenges your thinking may operate at the same level of autonomy. But the experience, expectations, and way a person works with each can be very different.
We wanted a framework that captured both.
How the HIA Agent Scale works
The HIA Agent Scale classifies AI systems along two dimensions.
Responsibility describes how much of the work the AI handles independently, across six Levels:
Level 1: Chatbot
Level 2: Contextual Assistant
Level 3: Work Assistant
Level 4: Semi-Autonomous Agent
Level 5: Supervised Autonomous Agent
Level 6: Fully Autonomous Agent
Relationship describes the role the AI plays with the person using it, across five categories:
a: Answerer
b: Assistant
c: Thought Partner
d: Coach
e: Companion
Put the two together and you get a classification such as 4c: a Semi-Autonomous Agent operating as a Thought Partner.
The distinction is useful because two systems with similar technical capabilities may be designed for entirely different kinds of human interaction.
The HIA Agent Scale gives people a common vocabulary for that conversation. It helps clarify what a system is expected to do, where human involvement belongs, and what level of oversight its responsibilities require.
We’ve made the scale publicly available, including an interactive classification tool, examples, and definitions for each Level and relationship category.