Short answer
Artificial narrow intelligence (ANI), also called narrow AI or weak AI, is AI whose useful competence stays bounded to particular tasks, domains, tools, or deployment conditions. Under the common three-tier framework of narrow, general, and super intelligence, every AI system deployed today is narrow AI, including large language models and the coding agents built on them. Narrow does not mean simple: under this framework, a narrow system can match or beat humans on a specific task without showing broad, reliable transfer across most cognitive tasks.
As of September 5, 2026, there is no universally accepted determination that any deployed system has achieved human-level AGI. Under the common three-tier framework, deployed systems are classified as narrow AI; DeepMind's levels framework instead classifies some broad frontier models as Emerging AGI. These are framework-dependent classifications, not a field-wide consensus. A modern coding agent's task scope also depends on its agent harness, which supplies tools, context, permissions, and an execution loop.
- ANI is bounded competence: strong on particular tasks or domains, without broad, reliable transfer across most cognitive tasks.
- Under the common three-tier framework, deployed generative models and coding agents are classified as narrow AI.
- Narrow is not weak. A narrow system can be superhuman on one task and still be narrow AI.
- Agentic behavior comes largely from the agent harness around a narrow model, not from general intelligence.
What artificial narrow intelligence means
Artificial narrow intelligence is the category for AI that performs within a bounded scope. The boundary can be a task, such as translation or code completion, a domain, such as radiology images, or a set of tools and deployment conditions the system was built and permitted to operate within. The classic three-tier teaching framework places ANI below artificial general intelligence (AGI) and artificial superintelligence (ASI).
The phrase weak AI is often used interchangeably with narrow AI, though in philosophy weak AI can also carry a separate meaning about whether a system genuinely understands anything. For a capability discussion, treat ANI, narrow AI, and weak AI as the same tier: competence without general transfer.
Why narrow does not mean simple or weak
The most common error is reading narrow as limited-in-power. A narrow system can be dramatically superhuman inside its scope. A game engine can beat every human player alive and still be narrow AI, because that skill does not transfer to writing an essay or diagnosing a fault. Narrowness is about the breadth of transfer, not the ceiling of performance.
This is why capability and category are separate questions. Asking how good a system is at its task is different from asking whether its competence generalizes across domains the way a human's does.
How narrow AI works
Most modern narrow AI learns a mapping from inputs to outputs from data, then applies that mapping within the conditions it was trained and deployed for. A large language model predicts text; a vision model classifies images; a recommender ranks items. Each is optimized against an objective inside a scope, and each degrades or fails outside that scope rather than adapting the way a general intelligence would.
Examples of ANI in 2026
Narrow AI is nearly all of the AI in production today. A representative spread:
- Recommendation and ranking systems that order feeds, results, and products.
- Computer vision and speech recognition for images, video, and audio.
- Generative AI: text, image, audio, and code models that produce content within their trained modalities.
- Coding agents and other tool-using agents that plan and act across multiple steps inside a bounded environment.
Why today's coding agents still fit the narrow tier
A modern coding agent can work across many files, several languages, and long multi-step tasks. That breadth is real, and it is why people ask whether such agents have crossed into general intelligence. Under the standard three-tier framework they have not: their competence is still bounded by the digital environment they run in, the tools they are given, the permissions they hold, their reliability, and the deployment scope their operator set.
Narrow, in this sense, does not mean the agent can only output one kind of thing. It means the useful competence does not generalize, on its own, beyond the conditions engineered around it. One honest caveat: some research frameworks, such as DeepMind's levels of AGI, classify broad-but-unreliable frontier models as emerging AGI rather than narrow, so the tier boundary is a framework choice, not a settled fact.
Model vs agent harness: where the capability comes from
When a model appears to act broadly, the surrounding system can extend what it accomplishes by supplying context, tools, permissions, and an execution loop. The model contributes reasoning and generation; the harness turns those outputs into actions and completed work. This is also why the same model can feel brilliant in one product and limited in another.
- Model
- Reasoning, generation, and interpretation.
- Instructions and context
- Task framing and the relevant information for this run.
- Tools
- File, shell, browser, API, and domain actions the agent can take.
- Execution loop
- Plan, act, observe, and retry until done or stopped.
- Permissions and sandbox
- The bounded action space the run is allowed.
- Recovery and evaluation
- Error handling and completion checks.
An agent harness can draw broad, useful behavior out of a bounded model by giving it context, tools, memory, and an execution loop. That makes the system more agentic. It does not, by itself, establish general intelligence.
ANI vs generative AI, agentic AI, AGI, and ASI
- Generative AI describes systems that generate content; today's generative models are generally classified as narrow under the common three-tier framework.
- Agentic AI describes how autonomously a system acts, not which capability tier it occupies; an agent can be highly autonomous and still be narrow.
- AGI would show broad, roughly human-level competence with transfer across most cognitive tasks.
- ASI would exceed the best human performance across virtually all important domains.
How to evaluate narrow AI in production
Because narrow AI is defined by its scope, evaluation should be scoped too. Judge a system on the task you actually need, with an explicit pass criterion, rather than on a general intelligence label. For agents, that means measuring completion, cost per completed task, and reliability on representative work, which is the approach behind model evals and same-task benchmarking.
Narrow models, broad-looking systems, and the harness layer
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.
The practical lesson of the narrow tier is that today's most capable-looking agents are narrow models wrapped in a capable harness. Treating the harness as a first-class, swappable component is what lets a bounded model do broad, useful work, without pretending the underlying capability tier has changed.
FAQ
- What is ANI in simple terms?
- ANI, or artificial narrow intelligence, is AI whose useful competence remains bounded to particular tasks or domains. It lacks broad, reliable transfer across most cognitive tasks. Under the common three-tier framework, today's deployed systems are classified as ANI; other taxonomies may label some frontier models Emerging AGI.
- Is generative AI a type of narrow AI?
- Today's generative AI systems are generally classified as narrow under the common three-tier framework because their useful competence remains bounded, although DeepMind's levels framework classifies some broad frontier models as Emerging AGI.
- Are ChatGPT and coding agents ANI?
- Under the common three-tier framework, yes: they are narrow AI. Some research frameworks classify broad-but-unreliable models as emerging AGI instead, so the label depends on the framework, but neither treats them as general intelligence in the full sense.
- What is the difference between narrow AI and weak AI?
- They are usually used interchangeably for the same tier. In philosophy, weak AI can carry an extra meaning about whether a system truly understands anything, but for a capability discussion narrow AI and weak AI mean the same thing.
- Can ANI outperform humans?
- Yes. A narrow system can be superhuman on its specific task and still be narrow AI, because the skill does not generalize. Narrowness is about breadth of transfer, not the performance ceiling.
See the anatomy of an agent harness
Understand how an execution loop, tools, context, and permissions turn a narrow model into a system that completes real work.
Read the agent harness guide