The six dimensions of AI agents: how to measure AI agenticness

Agentic AI exists on a spectrum of complexity. This spectrum can be measured across six distinct capabilities, from assistive tools right through to completely autonomous agents. Measuring AI agenticness will allow you to better manage the risks associated with the AI systems, establish appropriate ethical guardrails and achieve a strategic advantage from scaling agentic AI.

AI agenticness: the six dimensions of the spectrum

Agentic AI differs from other AI tools thanks to two primary capabilities: its ability to perform complex tasks and to operate independently. On top of these primary capabilities, four secondary capabilities (generality, adaptability, environmental complexity and impact) further influence the system’s level of agenticness. Together, these capabilities can be described as the six dimensions of AI agenticness spectrum.

An AI system that is highly developed across these six dimensions is more agentic and therefore more impactful than a system that is less developed.

Primary capabilities of agentic AI systems

The two primary agentic measurements – goal complexity and independent execution – are important determinants on whether an AI system crosses the line into being agentic. 

Goal complexity

The system is able to break down, balance, sequence and achieve complex tasks. An agentic AI system with low-level goal complexity pursues a single goal in a logical, self-determined manner — think of an automated notetaker capturing the highlights of a meeting. In contrast, a system with high goal complexity can deconstruct intricate challenges into interdependent subgoals where effective sequencing is critical. For example, an intelligent IT monitoring system can check a complex cloud environment for potential problems, diagnosing the cause of outages and carrying out repairs.

Independent execution

The system is able to execute tasks and achieve goals with limited human intervention. A low level of independent execution means the system performs routine tasks while operating within a narrow domain — for example, issuing a refund to a customer for a damaged product. High independent execution means the system can plan, sequence tasks, execute subroutines and initiate actions without human intervention based on high-level goals, learning continuously and autonomously coordinating with other agentic systems. Think of an autonomous cybersecurity triage system that detects and responds to cyber threats in real time.

Secondary capabilities of agentic AI systems

The four secondary capabilities are useful in determining the level of agenticness of the system.

Generality

The system can operate across multiple domains, with its level of generality varying according to its success at operating across those domains. A customer service agent that only handles billing questions has low generality. On the other hand, a system that can seamlessly handle billing, technical support and account management has higher generality.

Adaptability

A system with a high level of adaptability can learn from experience, adjust its actions in real time and shift execution strategies when it encounters unexpected scenarios or unforeseen obstacles. As an example, an adaptable coding agent can write code, execute tests, interpret error messages and rewrite the code.

Ability to handle environmental complexity

The system can cope with environmental unpredictability and complexity. Simple environments tend to be predictable, with clear feedback loops. On the other hand, complex environments tend to involve multiple stakeholders, competing signals, incomplete information and cascading consequences. A trading system operating in volatile markets with geopolitical uncertainty faces higher environmental complexity than one managing stable bond portfolios.

Impact

The actions of the AI system have an impact on the business and potentially on wider society – with the scale of this impact influencing the overall agenticness of the system. A typo correction tool typically has minimal impact. On the other hand, a drug discovery tool or an autonomous vehicle could profoundly impact an organization’s business model, as well as society more broadly.

AI agenticness: low, medium and high

The following examples of agentic systems, that can be classified as low, medium and high on the agenticness spectrum.

How to measure AI agenticness

By measuring the six dimensions of your agentic AI system – both its primary and its secondary capabilities and its proficiency in those capabilities – you can assess whether the system is working safely and reliably and in line with human objectives. You can also identify possible performance improvements.

To measure the agenticness of your AI system, answer the following questions, ranking the system’s capability as low, medium or high:

• Can your system identify and act on goals without a rigid, predefined formula?
• Can your system achieve objectives without you specifying every step?
• Can your system operate successfully across different business domains – for example, combining billing advice with technical support?
• To what extent is your system capable of recalibrating its approach when conditions change unexpectedly?
• Can your system cope with a chaotic, unpredictable environment where the goalposts are continuously shifting?
• What are the intended outcomes and potential unintended consequences of your system’s capabilities?

Key takeaways for leaders on AI agenticness

• Agenticness is a six-dimensional spectrum, not a yes or no question
• To be considered agentic, systems must have the primary capabilities of agentic AI: goal complexity and independent execution
• Combined with its primary capabilities, an AI system’s level of agenticness will be determined by its secondary capabilities – generality, adaptability, environmental complexity and impact
• The sophistication of governance around an AI system must scale in line with its agenticness