Short answer
Artificial general intelligence (AGI) generally refers to AI with broad, roughly human-level capability across cognitive tasks, including the ability to transfer knowledge and adapt to genuinely new tasks. It sits above narrow AI, which is bounded to specific tasks, and below artificial superintelligence, which would exceed the best humans across virtually all domains. There is no single, universally accepted operational definition of AGI; different groups anchor it to human-level generality, to economically valuable work, or to measured breadth and reliability.
As of September 5, 2026, there is no universally accepted determination that a deployed system has achieved AGI. Some company leaders have claimed to be at or past it, but such claims are individual or corporate judgments, not a scientific consensus, because the field lacks an agreed test. A capable AI agent is not automatically AGI: tools and an agent harness can expand task scope and autonomy without by themselves demonstrating general intelligence.
- AGI means broad, roughly human-level competence with transfer across most cognitive tasks, not mastery of one task.
- There is no single accepted definition or test, so treat specific AGI claims as framework-dependent, not settled fact.
- As of 2026, no deployed system is agreed to be AGI; leader claims are judgments, not consensus.
- An agent being autonomous or agentic does not make it AGI; autonomy and capability breadth are measured separately.
What artificial general intelligence means
AGI is the middle tier of the common three-level framework: narrow AI, general AI, and superintelligence. Where a narrow system is competent inside a bounded scope, a general intelligence would handle a wide range of cognitive tasks at roughly human level, carry knowledge from one domain into another, and adapt to problems it was not specifically built for.
The emphasis is on breadth and transfer. A system that is superb at one task, even superhuman at it, is still narrow. Generality is the property that a single system can take on unfamiliar tasks across domains without being rebuilt for each one.
Why there is no single accepted definition of AGI
Different serious groups define AGI differently, which is why headlines conflict. Three common anchors:
- Human-level generality: matching a typical human across most cognitive tasks. OpenAI's charter, for instance, frames AGI as highly autonomous systems that outperform humans at most economically valuable work, an institutional definition rather than an industry-wide one.
- Economically valuable work: the ability to do the bulk of what humans are paid to do, which ties the definition to labor rather than to any single benchmark.
- Performance, breadth, and reliability: measuring generality and competence as separate axes, the approach DeepMind's levels of AGI framework takes.
What capabilities would an AGI need?
Across definitions, a few capabilities recur as necessary for something to count as general:
- Knowledge transfer: applying what it learned in one domain to a different one.
- Learning new tasks: acquiring competence at problems it was not specifically trained for.
- Reasoning across domains: connecting facts and methods from unrelated areas.
- Reliable performance in novel situations: not just occasional success, but dependable competence outside the training distribution.
- Knowing when to ask for help: recognizing the limits of its own competence.
AGI vs AI, generative AI, agentic AI, and ASI
| Dimension | Term | What it actually refers to |
|---|---|---|
| Narrow AI | Narrow AI (ANI) | Bounded competence on specific tasks or domains. All AI in use today. |
| Generative AI | Generative AI | Describes systems that generate content; today's generative models are generally classified as narrow under the common three-tier framework. |
| Agentic AI | Agentic AI | A description of how autonomously a system acts, not a capability tier. An agent can be narrow. |
| AGI | AGI | Broad, roughly human-level competence with transfer across most cognitive tasks. |
| ASI | ASI | Competence beyond the best humans across virtually all important domains. |
Is today's AI already AGI?
Frontier models are broad in a way earlier AI was not, which is what makes the question live. Under the standard three-tier framework the mainstream answer is no: current systems are narrow AI that is unusually broad, not general intelligence, because their reliability and transfer fall short of the human baseline the definitions require.
The honest complication is that this depends on the framework. DeepMind's levels of AGI, for example, would place broad-but-unreliable frontier models in an emerging AGI band rather than in the narrow tier. So the truthful statement is not AGI does not exist but rather there is no agreed test, and under the common framework no deployed system is considered general intelligence as of 2026.
Why an AI agent is not automatically AGI
It is easy to mistake a capable agent for a general intelligence because a good agent visibly plans, uses tools, edits code, runs tests, and recovers from failures. Those behaviors can arise from a combination of model capabilities and deployment scaffolding, and they do not by themselves establish domain-general cognition.
Walk through one coding task: the model decides the next step; the agent harness supplies the execution loop, repository context, shell, file permissions, sandbox, tests, and retries; and the combination inspects the codebase, edits several files, runs the tests, and fixes what broke. The result is more agency, inside a bounded set of tools, an environment, permissions, and a task scope. Agentic behavior does not establish general intelligence.
How agent harnesses produce general-looking behavior
HarnessRouter is the world's first unified interface for agent harnesses. A unified interface means one API contract for running complete agent harnesses: starting tasks, streaming progress, continuing sessions, and collecting files and results work the same way across every harness on the platform.
From a production perspective, a harness changes how a model is deployed: it frames tasks, grants tools, preserves context, and manages execution. These features can increase task completion and autonomy, but neither autonomy nor scaffolding alone establishes the breadth and performance required for AGI. The model-and-system combination still has to be evaluated across a broad task set.
How could AGI be measured?
If there is no single definition, measurement is unavoidably plural. A few points of agreement:
- The Turing test is not sufficient: fooling a human in conversation does not establish broad, reliable competence.
- DeepMind's levels of AGI framework separates generality from performance and treats capability and autonomy as distinct axes.
- Real-world task evaluation matters more than any single quiz: measuring completion, cost, and reliability on representative work, in the spirit of model evals.
Potential uses and risks of AGI
Because AGI does not yet exist by consensus, its uses and risks are discussed as possibilities, not facts. Potential benefits include broad problem-solving across science, engineering, and services. Potential risks include misuse, loss of control, and concentration of power. The responsible way to discuss both is to keep predicted timelines and outcomes labeled as forecasts rather than stating them as established results.
FAQ
- What is AGI in simple terms?
- AGI is AI that could handle a broad range of cognitive tasks at roughly human level and carry what it learns from one area into another. That breadth and transfer is what separates it from today's narrow AI, which is strong only within a bounded scope.
- Has AGI been achieved?
- As of September 5, 2026, there is no universally accepted determination that any deployed system has achieved AGI. Some leaders have claimed to be at or past it, but those are individual or corporate judgments, and the field has no agreed test to settle the question.
- Is ChatGPT an AGI?
- Under the common three-tier framework, no: models like ChatGPT are broad narrow AI, not general intelligence. Some research frameworks would call broad frontier models emerging AGI, so the answer depends on the framework, but neither treats them as full general intelligence.
- Is agentic AI the same as AGI?
- No. Agentic describes how autonomously a system acts, not how broad or strong its intelligence is. A system can operate with substantial autonomy while remaining narrow; capability breadth, performance, and autonomy must be evaluated separately.
- Does AGI require consciousness or self-awareness?
- Not under mainstream capability definitions. AGI is generally defined by what a system can do across tasks, not by whether it is conscious. Consciousness is a separate philosophical question that the common definitions do not require.
See how a harness turns reasoning into work
Understand how an execution loop, tools, and context turn a narrow model's reasoning into completed tasks.
Read the agent harness guide