The Engineering Definition of AGI

AGI has long been regarded as a central goal of artificial intelligence, but it has never had a stable, operational definition. Many definitions emphasize being “as intelligent as humans,” “able to perform a wide range of tasks,” “conscious,” or “able to outperform humans in economically valuable work.” Each of these formulations has historical value, but they struggle to answer a key question: how can we tell whether AGI is getting closer?

This article proposes an engineering definition:

AGI means AI can autonomously complete complex projects that originally required long-term collaboration by ordinary human teams.

More formally:

Engineering AGI refers to an AI system that, after goals, resources, safety boundaries, and initial context are provided, can stably complete most typical complex projects that would originally require long-term collaboration by ordinary human teams, without continuous human driving, and can assume verifiable delivery responsibility for the project result. “Complete” does not require the AI to personally perform all physical operations, real-world authorization, or high-risk ethical judgment; it means the core cognitive work, process governance, and delivery judgment required to advance the project are primarily handled by AI.

The core issue is not whether AI resembles humans, but whether the main cognitive work, process governance, and delivery judgment required to advance a project can be stably handled by AI.

1. A Good AGI Definition Should Satisfy Three Conditions

An AGI definition with practical guidance value cannot be only a philosophical imagination, nor can it use outcomes to describe outcomes. It should satisfy at least three conditions.

First, it should be logically falsifiable.
If a system cannot reliably take on complex projects, we should be able to say clearly: it is not AGI yet. A definition that cannot be falsified easily degenerates into belief or marketing language.

Second, improvement should be acceptable.
If one system is closer to AGI than a previous generation, we should be able to explain where it has improved: whether it reduces human intervention, expands the scale of projects it can take on, improves execution stability, or strengthens self-correction.

Third, progress should be measurable.
AGI should not be judged only by point tests. It should be evaluated through metrics such as project duration, task dependency depth, time without human intervention, delivery quality, number of human takeovers, maintenance stability, and cost efficiency.

Only when these conditions are met can AGI move from a vague concept to an approachable engineering goal.

2. Complex Project Loops Are What Correspond to a Productivity Leap

AGI matters not because it satisfies a certain imagination of intelligence, but because it should constitute a turning point in human society’s productivity.

Most key forms of human productivity are connected to the ability to build complex projects. Software systems, drug development, manufacturing systems, energy systems, education platforms, scientific engineering, infrastructure, and organizational operations are not point tasks in essence. They are long-term collaborative engineering efforts.

Current AI has already significantly improved the efficiency of point tasks. It can write code, generate documents, analyze materials, answer questions, and assist design. But in most real scenarios, AI is still a human-driven tool. Humans propose the goals, humans decompose the tasks, humans define the boundaries, humans accept the results, and humans take final responsibility.

The turning point of engineering AGI is this: AI is no longer merely a capability module, but begins to take project responsibility. Humans shift from continuous drivers to goal setters, boundary setters, auditors, and final authorizers.

This is the core of the productivity leap. Once the execution cost of complex projects falls, many projects that previously could not be started because organizational cost, coordination cost, trial-and-error cost, and maintenance cost were too high will become possible.

3. Engineering AGI Does Not Require Complete Physical-World Interaction

A common misunderstanding is that if AI cannot directly operate the physical world like a human, it cannot count as AGI. This requirement is unreasonable.

In reality, complex projects are not completed by one person performing every physical operation. A project manager does not personally manufacture every component, an architect does not personally execute every installation step, and a company leader does not personally complete every on-site task. Complex projects rely on goal systems, collaboration systems, tool systems, and execution interfaces.

Therefore, the key to engineering AGI is not whether AI has a humanoid body, but whether it can complete project loops through available interfaces. These interfaces may be software APIs, robots, automated equipment, sensor systems, specialized tools, or ordinary people performing physical operations under AI guidance.

Humanoid robots and specialized automation equipment will expand the range of projects AGI can take on, but they are not prerequisites for the definition. AGI does not need to first become a complete human before it is qualified to change productivity.

4. Older AGI Definitions No Longer Fit the Current AI Reality

AGI definitions are not necessarily absolutely right or wrong. At different historical stages, people’s imagination of AGI often comes from the most prominent form of AI capability at the time.

Early AI was far from stable natural language understanding, reasoning, and knowledge integration, so AGI definitions naturally emphasized the “feeling of intelligence”: whether it seemed to think like a person, whether it could converse naturally, whether it could answer a wide range of questions, and whether it approached or surpassed ordinary people on local cognitive tasks.

But today’s strongest language models have already surpassed ordinary people on many local tasks. If AGI only means “reaching or exceeding ordinary people on many local intellectual tasks,” then AGI in that sense has already been achieved to a considerable degree.

The problem is that this has not directly produced a productivity leap.

The real bottleneck is not local intelligence, but complex project capability: task decomposition, reliable delivery, autonomous acceptance, long-term execution without drift, effective correction after failure, cross-context goal consistency, and continuous maintenance and improvement.

Therefore, AGI definitions that emphasize conversational intelligence, humanlike feeling, local task performance, and comparison with ordinary human IQ are increasingly unsuitable for the current moment. They were once illuminating, but they cannot explain what AI still lacks before it can produce a true productivity leap.

5. Why the Definition of AGI Matters So Much

The definition of AGI matters not because we need to compete over a word, but because AGI is the most widely circulated imagination of AI’s forward path in contemporary society.

Today’s large language models already possess strong local intelligence, yet most people still cannot clearly explain one phenomenon: why is AI already so smart, while a broad productivity leap has still not fully occurred? Why have ordinary repetitive jobs not disappeared at scale? Why have enterprise organizations not been thoroughly restructured?

If AGI is defined as “being as smart as humans” or “exceeding ordinary people on local tasks,” then current AI already seems very close to AGI. But that definition cannot explain reality. What truly limits the productivity leap is complex project loop capability.

A good AGI definition can unify the AI narrative at the lowest cost. It lets the public understand that AI’s core progress is not just becoming smarter inside a chat box. It lets industry understand that what is truly worth tracking in the next stage is not only exam scores, question-answering ability, code snippet quality, or multimodal demos, but reliable delivery, long-term maintenance, failure correction, and autonomous closed loops.

Therefore, the definition of AGI is not a distant philosophical issue. It is public narrative infrastructure for the AI era. If the definition is wrong, the industry will repeatedly overestimate the social impact of local intelligence and underestimate the bottleneck of complex project governance. If the definition is clear, AI’s development path becomes understandable, assessable, and discussable.

6. The Shortcomings of Existing AGI Definitions

Existing AGI definitions can be roughly divided into several categories. They all have historical value, but most cannot satisfy the engineering requirements of falsifiability, acceptability, and measurability.

The first category is Turing-test-style definitions.
The representative source is Alan Turing’s 1950 paper “Computing Machinery and Intelligence.”¹ It asks whether a machine can make people unable to distinguish it from a human in conversation. This direction contributed by shifting the judgment of intelligence toward external behavior, but it tests anthropomorphic performance rather than complex project capability. A system can be highly humanlike in conversation while being unable to maintain real engineering work over time.

The second category is strong AI or consciousness definitions.
The representative source is John Searle’s 1980 paper “Minds, Brains, and Programs.”² These definitions ask whether machines truly possess mind, understanding, or subjective experience. This is an important philosophical question, but it is hard to turn into an engineering acceptance standard. Even if a system can reliably complete complex projects, we still may not be able to prove whether it is conscious; conversely, even if some form of consciousness exists, that does not mean it has complex project capability.

The third category is human-level intelligence definitions.
A representative source is the definition of High-Level Machine Intelligence used by Grace et al. in “When Will AI Exceed Human Performance? Evidence from AI Experts.”³ These definitions try to compare AGI with human labor capability, but the boundary of “human-level” remains vague: ordinary people, experts, or teams? More importantly, complex achievements in human society are not completed by isolated individuals, but by teams, tools, processes, and institutions.

The fourth category is economically valuable work definitions.
The representative source is the OpenAI Charter’s description of AGI as highly autonomous systems that outperform humans at most economically valuable work.⁴ This type of definition connects AGI to productivity, making it closer to reality than “being as smart as humans.” But “most economically valuable work” is still too macro and easily becomes a retrospective judgment. It describes the possible result of AGI, but is not enough to guide how progress should be accepted along the way.

The fifth category is formal general intelligence definitions.
The representative source is Legg and Hutter’s 2007 paper “Universal Intelligence: A Definition of Machine Intelligence.”⁵ These definitions try to describe, in theoretical terms, an agent’s ability to achieve goals across a wide range of environments. They are highly abstract and more rigorous, but far from how AGI is judged in real industry. Engineering AGI needs reproducible project evidence, not only a purely formal description.

The sixth category is AGI level frameworks.
The representative source is Google DeepMind’s “Levels of AGI for Operationalizing Progress on the Path to AGI.”⁶ Such frameworks try to describe AGI progress by capability breadth, performance depth, and degree of autonomy, which is better than a single endpoint definition. But most level frameworks still emphasize task capability rather than project loops. Strong task capability does not mean the system can take responsibility continuously throughout a long-term complex project.

The seventh category is social-impact definitions.
The representative source is Open Philanthropy’s definition of Transformative AI: AI that could trigger a transition comparable to, or more significant than, the agricultural or industrial revolution.⁷ This is suitable for macro discussion, but not as a capability definition. Social impact can often only be judged after the fact and is affected by institutions, costs, regulation, and deployment speed. It can describe the consequences of AGI or advanced AI, but it cannot replace the standard for judging AGI.

The problem with these definitions can be summarized in one point: they are either too much like philosophical judgments, too much like retrospective outcomes, or stuck at local task capability. They do not place the unit of AGI judgment where real productivity matters most: the complex project loop.

7. The Boundary of This Definition

To avoid overclaiming, this definition should be publicly expressed together with its boundaries.

First, it is not the final definition of AGI in every possible sense. The definition in this article is better suited for discussions of software development, AI engineering, knowledge-work automation, complex project delivery, organizational production substitution, and investment or industry judgment. It places the unit of judgment on complex project loop capability, so it is especially useful for analyzing when AI can create a real productivity leap.

But it does not directly cover every AGI discussion scenario. For example, questions of consciousness, subjective experience, embodied robotics, open-world survival, the complete process of scientific discovery, the full capabilities of military systems, and home care or emotional companionship should not be directly subsumed under this definition. They may be related to AGI, but they are not the core problem this definition is trying to solve.

Second, “autonomous” does not mean no human gives the goal. Autonomy here does not mean that no one proposes the goal, provides resources, sets safety boundaries, or that humans completely exit the responsibility chain.

In this article, autonomy means that once the goal, boundaries, toolchain, and acceptance principles are given, the AI system no longer needs humans to continuously intervene in task decomposition, execution repair, context recovery, and process governance.

Therefore, this definition does not remove human goal setting, boundary setting, resource authorization, risk auditing, or final responsibility. It emphasizes whether AI can, after receiving a goal and boundaries, independently take on the continuous execution responsibility inside a complex project loop.

8. The Engineering Definition Turns AGI Into an Approachable Goal

The value of the engineering definition of AGI is that it turns AGI from a vague imagination into an approachable goal.

We no longer need to repeatedly argue about whether AI is “like humans.” Instead, we can ask more concrete questions:

• How long a project cycle can it take on independently?
• How deep a task dependency chain can it handle?
• How much human intervention does it require?
• Can it autonomously propose acceptance criteria?
• Can it discover and fix its own errors?
• Can it re-plan when requirements change?
• Can it maintain quality during long-term maintenance?
• Can it complete closed loops across software, data, documents, and physical-world interfaces?

These questions can be tested, compared, and accumulated as evidence. They can also improve gradually as models, tools, and engineering systems improve.

Therefore, engineering AGI does not downgrade AGI into a software engineering problem. It puts AGI back where human productivity actually happens. Human civilization does not advance through point intelligence alone, but through complex project capability. The true sign of AGI should not be how smart it appears in a particular question-and-answer exchange, but whether it can reliably take over the construction, maintenance, and evolution of complex projects.

When AI can autonomously complete complex projects that originally required long-term collaboration by ordinary human teams, AGI truly moves from concept to reality.

---

References and Representative Sources

Footnotes

1. Alan M. Turing, Computing Machinery and Intelligence, Mind, 1950. ↩

2. John R. Searle, Minds, Brains, and Programs, Behavioral and Brain Sciences, 1980. ↩

3. Katja Grace, John Salvatier, Allan Dafoe, Baobao Zhang, Owain Evans, When Will AI Exceed Human Performance? Evidence from AI Experts, arXiv, 2017; JAIR, 2018. ↩

4. OpenAI, OpenAI Charter, 2018. ↩

5. Shane Legg, Marcus Hutter, Universal Intelligence: A Definition of Machine Intelligence, Minds and Machines, 2007. ↩

6. Meredith Ringel Morris et al., Levels of AGI for Operationalizing Progress on the Path to AGI, arXiv, 2023; ICML, 2024. ↩

7. Open Philanthropy, What Open Philanthropy Means by “Transformative AI”, 2019. ↩