AGI Dimensions: How to Assess AI Risk Through Capability and Autonomy

🌏 閱讀中文版本

A PM forwarded you an article titled “The AGI Era Has Arrived.” The note said: “Should we fully automate this?”

You stare at the CI/CD pipeline. You’re not thinking about automation. You’re thinking: “If it deletes the database, who is responsible for recovery?”

The answer is usually: you will be the one sweating a little more tomorrow morning.

Treating an AI project as a single black box is one of the most common blind spots in engineering evaluation. You think you’re evaluating a whole species. In practice, you’re only looking at one dimension pushed to its limit.

In The Levels of AGI Framework, Google DeepMind proposed a classification system. Beyond Performance Levels and Generality, the paper makes another key dimension explicit: Autonomy Levels. It measures one thing: how much human oversight does the AI need while doing the task?

When teams mix up these dimensions, they usually drift into two decision biases:

1. Overly cautious deployment: they assume the system already has independent agency, so they avoid deploying any automation at all.
2. Premature delegation: they see strong capability, then let it act alone in critical decisions.

What we need is a matrix, not a label.

Capability Levels: From “Emerging” to “Superhuman”

One of the paper’s core contributions is breaking AGI into Performance Levels. This is not about whether it is “human.” It is about who it performs better than on cognitive tasks.

Keep these six levels in mind. They are progressive, not jump cuts:

Based on the paper’s framework and current public capabilities, as of July 2026, state-of-the-art models are generally classified as L1 (Emerging AGI).

They perform very well in coding, translation, and summarization, which shows competence. But in logical reasoning, long-horizon planning, and complex mathematical proof, performance is still unstable, which keeps them in the emergent range.

The word “emergent” is risky. It suggests that if you scale enough, competence will simply appear. That is not necessarily how this works. Moving from L1 to L2 may require improvements across model architecture, tool use, evaluation, and interaction design, not just more data piled on top.

Autonomy Levels: From “Tool” to “Agent”

To evaluate deployment risk, we need the paper’s other core dimension: Autonomy Levels. It measures one thing: how much human oversight does the AI need while doing the task?

The word “agent” means different things in different product contexts. From an engineering perspective, it is more useful to return to this permission boundary.

Pay attention to the L4 definition: “Autonomous execution… but knows when to consult humans.” This is not “no supervision.” It is “dynamic supervision.”

Most current API calls and chatbot replies sit between L0 and L2. You give the instruction. It gives the output. You decide what happens next.

The risk appears when teams try to place L1 capability directly into an L4 setting. It is like giving a root permission set to an L1 script. It can run. But once it makes a mistake, the whole system can unravel.

Orthogonality: Why These Two Dimensions Must Stay Separate

This is the paper’s most important insight: Performance and Autonomy are orthogonal.

That means you can have a high-capability, low-autonomy system. You can also have a low-capability, high-autonomy system. Most risk assessment errors come from mixing up the two.

Think of capability as engine horsepower. Think of autonomy as the steering wheel.

Put a race engine behind a toy steering wheel, and it drives straight into impact. Put a toy engine behind a race steering wheel, and it flips in the turn. The two need to match. This is not about right or wrong. It is about fit.

Case 1: Low to Mid Capability, Low Autonomy (L1 Performance, L1 Autonomy)

Scenario: customer support bot.
Capability: L1. It understands 90% of common questions and gives accurate answers.
Autonomy: L1. It proposes a response, but a human has to approve it before sending.

Risk: low. Even if it gives the wrong answer, a human support rep can intercept it. The marginal impact of the mistake stays at the information layer.

Case 2: Low Capability, High Autonomy (L1 Performance, L4 Autonomy)

Scenario: automated code deployment script.
Capability: L1. The code it generates may contain logic errors or skip edge cases.
Autonomy: L4. It runs deployment automatically, and humans step in only after failure.

Risk: very high. Limited capability creates mistakes. High autonomy amplifies them quickly. This is what people mean by an automation disaster: the error gets accelerated, and you do not get much chance to stop it.

Case 3: High Capability, High Autonomy (L3 Performance, L4 Autonomy)

Scenario: medical diagnostic support system.
Capability: L3. On specific imaging tasks, it performs at an expert level.
Autonomy: L4. It can autonomously call diagnostic equipment, suggest follow-up tests, and create work orders for physician review, consulting a doctor only on difficult cases.

Risk: high. The capability may be strong enough to support autonomous decisions, but regulated environments add risk around misdiagnosis, process execution, accountability, and compliance auditability. In healthcare, you also need to account for regulatory compliance, human review, and audit logs.

The key point is this: you do not need L5 capability to deploy L4 autonomy. You only need enough capability to support the risk boundary created by that level of autonomy.

If a task has a very low error cost, for example re-entering a spreadsheet, then L1 capability plus L3 autonomy is a medium-risk setup and may be acceptable if the outcome is reversible. But if the error cost is “loss of funds,” then L1 capability plus L4 autonomy is high risk and becomes an unacceptable exposure.

Risk Matrix: How Do You Choose a Deployment Strategy?

Put these two dimensions on top of each other, and your risk map appears. Place your current AI project into it:

Practical Guidance

1. Evaluate capability first: don’t ask, “Is this model strong?” Ask, “Across 100 tasks, how many times does it make an irreversible mistake?”
2. Then decide autonomy: don’t ask, “Can we automate this?” Ask, “If it is wrong, who absorbs the cost, and can we afford that cost?”
3. Follow the matching principle: decide based on error cost, reversibility, and monitoring capacity. Do not treat L3 as a rigid threshold. If capability is L1, keep autonomy at L2 or below. If you need L4 autonomy, make sure capability is at least L3, and put a strong rollback mechanism in place.

Market Signal and What Comes Next

Current state-of-the-art models are still in the L1 Performance stage. Under this paper’s classification, today’s models are closer to Emerging AGI than Competent AGI. If we use the capability and autonomy definitions in this article, claims that AGI has already arrived still need more shared benchmarks.

As model capability improves, we can expect to see:

• Broader L2 Performance: models will handle most cognitive tasks more consistently.
• Vertical pilots with expert-level performance: in finance, healthcare, and similar fields, we will likely see expert-grade autonomous assistive systems.
• Workflow pilots at L3-L4 Autonomy: in specific workflows, we will likely see expert-grade autonomous assistive systems.
• L4 Autonomy as a major debate: legal accountability and ethical frameworks around “AI agents” will become a primary discussion area.

The leap from L3 to L4 still lacks enough historical data. Looking back after N years, we may find that today’s framework was too optimistic.

Deploy AI in the right cell, not under a blurry headline.

Your Decision Checkpoint for Tomorrow

Before you press deploy, it is worth putting these three questions on the table:

1. What is the cost of error? Is it re-entering a form, or losing funds?
2. Who is responsible for recovery and cost absorption? If it is wrong, do you fix it, or can the system recover on its own?
3. What are you relying on? Are you relying on its capability, or on your monitoring?

Different answers lead to different strategies.

The market signal is not clear yet. But the risk boundary can be drawn now. Where will your ten minutes go tomorrow, drawing the boundary or handling the cost of exceptions?

Sources

• The Levels of AGI Framework — The original paper defining AGI capability, generality, and deployment autonomy discussions