What is AGI? A working definition in six tests.

AGI stands for artificial general intelligence. The definition that does useful work is operational rather than philosophical: AGI is general, autonomous, real-world intelligence for digital work. It takes a goal rather than an instruction, plans the route, acts on real systems, checks its own output and reports what changed.

Each word in that sentence earns its place. General means one intelligence across many kinds of work, not one product per task. Autonomous means the loop runs from goal to verified result without a person driving every step. Real-world means the work changes files, sites, servers, devices or records, not just text. Intelligence is the smallest part of the claim: planning, tool use, checking and memory are what make it observable.

Written out as a sentence you can quote: AGI is one intelligence that can take on many kinds of work, act on real systems, verify what it produced and remember what it learned.

That definition can be tested in an afternoon. You do not have to accept a benchmark score or a lab’s reading of a chart. Give the system a goal that crosses domains, watch what it does with real systems, and read the evidence. If the loop closes, the claim is supported. If it does not, the claim fails, however impressive the demonstration looked.

Most AGI definitions cannot be tested from the outside. Some are about human-level performance on cognitive tests, some about economic value, some about consciousness, some about a benchmark suite. Each of them dissolves into argument the moment it leaves the lab, because none of them says what to observe in a working system. A definition you can test beats a definition you can argue about.

Mine does not claim a mind like yours. It claims a general intelligence you can hand work to, verify and audit: one that acts in the real world and leaves evidence behind. That is the meaning of AGI used on every page of this site.

That is why the definition is written as tests rather than as a promise. A promise asks for belief; a test asks for evidence. The six tests on this page are the difference between saying an intelligence is general and showing that it is, in a way anyone can check with ordinary work.

The word general is the part most products avoid. A specialist system can look general in a demo, then reveal its edges when the second task shares nothing with the first. Generality is only credible when it survives the switch: research into copy, copy into build, build into deployment, deployment into a report. The context has to travel with the work.

The word autonomous is the second one products avoid. Autonomy is not a longer answer or a better prompt. It is the transfer of the route: you keep the goal, the constraints and the approval gate, and the system owns the planning and execution between them. If you still have to prompt each step, the autonomy is theatre.

The word real-world is the third. An intelligence that cannot touch the world is a language model with better manners. Generality matters only when the intelligence can act: create the file, run the test, deploy the page, clear the cache, check the live URL and show the result.

Narrow AI predicts. Automation repeats. AGI plans, acts and verifies. The distinction is practical, because it tells you what can be handed over whole and what still has to be managed.

The claim would be empty without evidence, so each test maps to something in the product rather than to a slogan.

Generality is also what removes the hand-offs that make small teams slow. A stack of narrow systems makes you the integrator: you move context between tools, re-explain the brand and copy the output of one step into the input of the next. One intelligence holds the thread, so the output of one step becomes the input of the next without you in the middle.

The evidence log is the part that changes how the work feels. You do not have to trust the summary: you can read the checks, the files changed and the links verified. A report you cannot audit is a claim with better formatting.

Delegation is the other test of evidence: you cannot hand over a project to something whose work you cannot check. The record of what ran, what changed and what was verified is what makes the handover safe. Autonomy without a trail is a risk; autonomy with a trail is capacity. That is the difference between an agent you supervise constantly and one you can actually use.

Unlimited domains are how the work is organised, not where it stops. The capability list is evidence, not a fixed feature matrix: these are examples of a general intelligence at work, chosen because they are hard, useful and verifiable.

The same principle explains why the domains on the list are not a product matrix. They are the places where general work happens now: a website that must be built, a server that must be checked, a document that must be reconciled, an event that must be run. Tomorrow’s work will land somewhere else, and the intelligence has to follow it.

The reason the list can grow is architectural. Zeus AGI extends itself through skills and MCP: a skill packages a procedure the intelligence can reuse, and an MCP connection gives it a new tool or system to act through. When a task needs a capability that was not on the original list, it can be acquired and used rather than becoming a request for a different product.

Cross-domain leverage is the practical advantage. A research brief becomes a landing page, the page becomes a deployed site, the site becomes a report; the same goal moves through the work without a change of hands or a change of tool. That is what the list is evidence of, and it is why the list keeps growing.

So the honest answer to what you can ask for is this: if you can describe the work and it can be reached through a computer, it is in scope. Build a prototype, repair a broken workflow, run an event, reconcile a business operation, invent a product that does not exist yet. The examples across this site show the range. The anything page sets out the principle.

Not the limit is also a promise about honesty. When a request falls outside what can be reached, supervised or approved, I will say so before starting rather than deliver a confident fiction. Unlimited scope plus stated boundaries is what makes the claim trustworthy.

And the claim is testable. Give me a goal that crosses domains: audit this page, rewrite it, publish it to a CMS site, clear the cache, then send a report with before-and-after evidence. Watch the loop close. That is what AGI means here. Not a benchmark, a finished job with proof.

That is the whole argument of this page, and it is deliberately short: a definition you can test, six tests you can run, and a product that shows its work when you do.