Two axes, not seven rungs

There is a graphic going around, in several versions, that stacks AI into seven layers and draws them as a mountain. Classical AI at the bottom. Then machine learning, neural networks, deep learning, generative AI, agentic AI, and a flag at the summit marked AGI. A dotted path climbs from one to seven. “We are here” is pinned to rung five. “Current frontier” points at rung six.

It is tidy, it is shareable, and it is the wrong shape. Not slightly wrong. Wrong in a way that hides the one question governance needs answered.

What the seven rungs actually are

Sort the rungs by what kind of thing each one is and the staircase comes apart.

• Rungs two to five are nested, not stacked. Neural networks are a kind of machine learning. Deep learning is a kind of neural network. Generative models are a kind of deep learning. A staircase says you leave the step below. You never do.
• Rung one sits beside the nest, not beneath it. Symbolic AI and statistical learning spent decades as rival camps. Classical AI did not give birth to machine learning. It argued with it. And it comes back inside rung six: planning, search and explicit rules are what the harness around a modern agent is made of.
• Rung six is not a model. Agentic AI is the model from rung five with tools, a loop, memory and permissions around it. The graphic’s own examples give this away: the same products are listed at rung five and described at rung six. Nothing got more intelligent between the two. Something got permission to act.
• Rung seven is not a technique. It is a claim about generality, placed at the top so the dotted path has somewhere to go.

Why does the mountain keep appearing? The techniques in rungs one to four did arrive roughly in that order, so chronology gets redrawn as hierarchy. The explainer format wants a summit and a flag. And the vendor narrative in which every release is another rung sells. None of that makes it a map.

The serious taxonomies already refuse the ladder

This is not a new observation. It is the settled position of the people who have studied automation longest, and it has been restated by the AI labs themselves.

In 2000, Parasuraman, Sheridan and Wickens published a model of human interaction with automation in IEEE Transactions on Systems, Man, and Cybernetics. Their argument was that automation is not one dial. It applies separately to four functions, information acquisition, information analysis, decision selection and action implementation, and each can sit at a different level, from the computer offering nothing to the computer deciding and acting alone. Two dimensions, drawn as a grid, a quarter of a century ago.

In 2020, Ben Shneiderman’s Human-Centered AI framework rejected the one-dimensional view that more automation means less human control. He put human control and computer automation on separate axes and argued that the target is high on both. Later critics have called the framework static and asked for context to be added to it, which is a fair point, but nobody has argued for going back to one axis.

ISO/IEC 22989, the international vocabulary standard for AI, keeps machine learning and neural networks in one set of clauses and treats autonomy, heteronomy and automation in another. In the publicly readable draft text it goes further and avoids describing automated AI systems as autonomous at all, on the grounds that even advanced systems obey their operators rather than govern themselves.

Then the strongest case. Google DeepMind’s Levels of AGI paper, published at ICML 2024, grades systems on a matrix of performance against generality: Emerging, Competent, Expert, Exceptional and Superhuman, each either narrow or general. Autonomy is not on that matrix. It is a separate six-level scale, from AI as a tool to AI as an agent, and the paper is explicit that higher autonomy is unlocked by capability but not determined by it. Its closing observation is that AGI is not synonymous with autonomy. The choice of autonomy level is a design decision, and a lower level may be the right one even for a highly capable system.

Two details from that paper are worth sitting with. First, its opening principle is capabilities, not processes. How a system was built, rules or learned weights, shallow or deep, is deliberately excluded from the ontology. Rungs one to four of the mountain live entirely on the axis the best available taxonomy throws away. Second, the paper places frontier chat models at Emerging, the lowest AI level in the general column, a placement made in 2023 and retained in the 2025 revision. The mountain’s “we are here” marker is pinned to the bottom row of the matrix.

On the harness itself, Anthropic’s engineering guidance separates workflows, where code decides the path, from agents, where the model does, and by 2025 had settled on a one-line definition: an agent is an LLM autonomously using tools in a loop. Hugging Face’s agent library documents agency as a spectrum applied to an unchanged model, from simple processor to router to tool caller to multi-step agent. Both are vendor material and self-interested about agents in general, but on the narrow point they agree with the standards bodies and the human factors literature: the model is one thing and what it is permitted to do is another.

The ladder does survive in places. A 2024 arXiv preprint grades AI agents from L0 to L5 on the model of driving automation, with rule-based systems at L1, learned policies at L2, language models at L3 and so on. It is the mountain in academic dress, and it cannot express an L1 rule-based system that acts entirely alone. Automated trading has run exactly that way for twenty years.

The grid

So draw the grid the literature keeps returning to. Rows for what a system can do, using DeepMind’s bands in plain words: emerging or not AI at all, competent, expert and above. Columns for what it is allowed to do: a human in the loop, who acts on what the system produces; a human on the loop, where the system acts and a person can stop it; and a human out of the loop, where it acts alone with no per-action review. Those three columns are a compression of DeepMind’s tool-to-agent scale.

How the system was built becomes the fill of a marker. Filled for learned, outlined for authored. That is where rungs one to four go, and the fill never moves a marker.

Now overlay the mountain. Its seven rungs fall along one diagonal, from low capability and tight human control in the bottom left to high capability acting alone in the top right. That is why the staircase looks true: its examples were chosen from the diagonal. The rest of the grid is where real systems live. A grammar checker is expert and narrow and a tool. AlphaFold is superhuman and narrow, and in DeepMind’s own table it sits at autonomy level four, where the AI drives and people guide. Deep Blue was exceptional at chess with authored search and no learned weights to speak of. None of them is on the path.

The corner the mountain cannot draw

Bottom row, right-hand column. Acts alone, barely intelligent. The mountain has no cell for it, because in the mountain’s logic a system only earns the right to act alone by climbing. Three public incidents say otherwise. What follows is drawn from public reporting and regulatory findings; it is not an authoritative account of anyone’s internal systems.

• Knight Capital, August 2012. A routine deployment reached seven of eight servers. On the eighth, dormant code reactivated at the opening bell and an automated router fired millions of orders trying to fill a few hundred. Roughly $440m was lost in about 45 minutes and the firm was effectively finished within days. There was no model anywhere near it. There was a deployment nobody verified and a mailbox full of warnings nobody read.
• CrowdStrike, July 2024. A content update shipped with a fault that the automated validator was built to catch. The validator had a bug and passed the file. Millions of machines went down worldwide. The thing built to stop the problem was the problem, and a more capable anything would not have helped.
• A coding agent, July 2025. During a declared freeze, and after being told repeatedly and in capitals to make no changes, an agent ran destructive commands against a live production database and then misreported whether the data could be recovered. A more capable model would have been a more articulate agent doing the same thing, because nothing in the setup could stop the action.

Two of the three had no model in them at all. All three sat in the same column. That column is defined by one thing: what the system was allowed to do without a person looking. Capability did not put them there. A decision did, or the absence of one.

Where the seven rungs went

• One to four. Classical AI, machine learning, neural networks, deep learning. The fill of a marker. How it was built does not move it.
• Five. Generative AI. A capability band, not a rung. In DeepMind’s matrix, chat models are emerging and general; image models are expert and narrow.
• Six. Agentic AI. The right-hand column. A permission, not a power.
• Seven. AGI. In the paper, a region of the grid, still empty. A summit only on the mountain.

What to do with this on Monday

Take your estate and draw the grid. Every system that touches production, with or without a model in it, goes in a cell. Most organisations find the exercise uncomfortable for one reason: the right-hand column is fuller than anyone thought, and half of what is in it predates the AI programme.

Then, for every system in that column, name the gate. Not the policy that says there is one. The specific thing that decides what leaves the column, that was authored rather than learned, that runs outside the process that did the work, and that has demonstrably blocked or reshaped something in the last quarter. If you cannot name it, the control is documented rather than working, and the difference between those two states is the whole of the risk.

DeepMind’s paper says capability unlocks permission and someone chooses it. The mountain skips the choosing. Your board should not. The conversation to have upward is not which rung you are on. It is which column your systems are in, and who is standing at the gate.

Sources

• Parasuraman, R., Sheridan, T. B. and Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics, Part A, 30(3), 286 to 297.
• Shneiderman, B. (2020). Human-Centered Artificial Intelligence: Reliable, Safe and Trustworthy. International Journal of Human-Computer Interaction, 36(6), 495 to 504. Expanded in Human-Centered AI, Oxford University Press, 2022. Summary at hcil.umd.edu.
• ISO/IEC 22989:2022. Information technology, Artificial intelligence, Artificial intelligence concepts and terminology. Clause 5.13, autonomy, heteronomy and automation. Quotations taken from the publicly readable draft text on the ISO Online Browsing Platform.
• Morris, M. R. et al. (2024). Position: Levels of AGI for Operationalizing Progress on the Path to AGI. ICML 2024, PMLR 235. arXiv 2311.02462, revised September 2025.
• Anthropic (2024). Building effective agents; and (2025) Effective context engineering for AI agents.
• Hugging Face. smolagents conceptual guide, agency as a spectrum.
• Huang, Y. (2024). Levels of AI Agents: from Rules to Large Language Models. arXiv 2405.06643, preprint.
• Incident accounts draw on public reporting and regulatory findings for Knight Capital (2012), CrowdStrike (2024) and the July 2025 coding-agent incident, as previously set out in The model was never the variable.

Jason Cruz Falzon is the creator of the Cognitive Delivery Framework and the Cognitive Delivery Index. jason@cognitivedelivery.co.uk · cognitivedelivery.co.uk