A Non-Technical Framework for AGI
Most AGI discourse is bad because it treats intelligence as binary. Either the model is smart or it isn’t. Either we’ve hit AGI or we haven’t.
A clearer way to think about it is through tiers of work. Instead of asking whether AI is generally intelligent, we should ask what kind of white collar labor it can perform.
There are three useful tiers:
- Assistant: executes tasks you specify
- Analyst: weighs evidence and tells you what it thinks
- Executive: sets direction, commits resources, and acts on decisions
Each tier has its own maturity curve. They do not unlock in sequence, and they fill in over time at different rates. To understand where we are on the path to AGI, we need to understand how far along each tier is.
My rough read is:
- Assistant: ~90% done
- Analyst: ~35%, entry level on a good day
- Executive: ~5%
The Assistant
Definition: the assistant executes clearly specified tasks.
Models can now summarize emails, rewrite paragraphs, extract data from documents, generate code, and turn vague instructions into usable output. This tier is close to solved.
When GPT first broke into the mainstream, models were mostly treated as fun chatbots. They hallucinated often, required careful prompting, and were easy to dismiss as toys. That has changed. Today, almost anyone can get useful output from a model.
More importantly, the prompting skill premium has collapsed. A power user and a first-time user now get roughly the same result from a given task. An extensive prompt about summarizing an article and “summarize this” get nearly the same result. You no longer have to prompt the model as if its life depends on the exact wording.
The assistant tier is not perfect, but for most everyday knowledge work tasks, it is already good enough.
The Analyst
Definition: the analyst does everything the assistant does, but adds judgment.
It can look at messy evidence, compare options, pressure test assumptions, and tell you what it thinks. An assistant helps you execute. An analyst helps you decide.
This is where current models become more uneven. For most users, models operate like an entry level analyst. They are capable of real analysis, but they are still unreliable, easily misled, and prone to confident wrongness when they make bad assumptions.
For power users, the analyst tier is closer to a senior level
analyst. Garry Tan’s GStack, for example, ships an
/office-hours skill modeled on real YC partner sessions.
It pressure tests your startup idea before you write a line of code.
Tan describes it as “only a 10% strength version of what a real
YC partner can do for you,” but still powerful enough to change
what people build.
The analyst tier is not yet a great human analyst, capable of replicating someone with 20+ years of experience. However, it is becoming a cheap, always available, competent analyst given the right context and prompting.
The remaining problem is less about raw model capability and more about calibration. The model can often do the work, but the system around it needs to know when to trust it, when to ask for more context, when to cite evidence, and when to say “I don’t know.”
For naive users to get senior analyst grade output, the harness matters as much as the model. The capability is emerging. The product layer is quickly catching up.
The Executive
Definition: the executive does everything the analyst does, but has authority to act.
This is the biggest jump in the three tiers. An analyst researches, pressure tests, and recommends, but the human still decides. An executive weighs tradeoffs, chooses a path, commits resources, and moves the work forward.
That distinction matters because the cost of failure changes. Analyst failures produce bad advice. Executive failures produce bad actions. The wrong email gets sent. The wrong money gets spent. The wrong customer promise gets made. The wrong product direction gets locked in.
We are only beginning to see executive level agents. Polsia’s tagline is “AI that runs your company while you sleep.” They pitch the product as an autonomous system that plans, codes, and markets your company 24/7. Whether their product is meeting expectations is almost beside the point. The category now exists, and people are buying into it.
The executive tier is not one giant AGI running a company. It is more likely to look like a small human team managing a swarm of assistant and analyst level agents. The human sets the direction, while AI systems increasingly handle the execution, analysis, and coordination beneath them.
That is why the executive tier is still early. It requires trust, delegation, accountability, and permission to create consequences in the real world. Models can already recommend actions. The hard part is giving them authority to take those actions safely.
Summary
So AGI will not arrive all at once. It will arrive unevenly, as more categories of labor move from human only, to human supervised, to machine executed. The assistant tier is nearly mature. The analyst tier is becoming real. The executive tier is just beginning.