Deep Dive: AI Learned From What We Created. Is It Now Learning From What We Do?

Something strange is happening in the AI mosh pit, for most of the generative AI era, the competition has appeared to revolve around intelligence. Technology companies have fought to demonstrate whose models can write more convincingly, reason more accurately, code more effectively and answer more difficult questions.

That description is rapidly becoming outdated. The frontier is moving from artificial intelligence that generates information towards artificial intelligence that takes action. Emerging systems can operate browsers, manipulate software, access files, communicate with external services and execute increasingly complicated sequences of work.

OpenClaw demonstrated the architectural significance of this transition by combining AI models with persistent memory, tools and access to a user’s computer. GPT-6 Astra, released by OpenAI in September 2026, advances another part of the technology stack by improving the underlying model’s ability to operate computers and complete professional workflows.

OpenAI reports that Astra scored 72.6 per cent on OSWorld 2.0, compared with 65.7 per cent for GPT-5.6 Sol, while completing the simulated tasks in approximately 47 per cent less time. It also recorded significant improvements in screen navigation, automation and terminal-based work.

However, computer control itself is not new. OpenClaw and other agent systems could already browse websites, execute commands, manage files and interact with applications. The more important development is the convergence of capable models with persistent systems that possess memory, permissions and access to the machinery of everyday life.

The central question is therefore no longer simply whether AI is becoming more intelligent. It is whether the AI movement is undergoing a fundamental mutation, or whether we are witnessing the latest stage of a much older technological pursuit which is gaining greater control over information, infrastructure, labour and the execution of human activity.

The First Internet Companies Wanted Our Attention

Lets take a step back the first great digital economy was built around attention. Search engines organised information, social networks organised relationships and smartphones created an interface that travelled everywhere with us. Recommendation algorithms increasingly determined what appeared on our screens, what attracted our interest and which version of reality reached our eyes.

The commercial prize was enormous because attention could be converted into advertising revenue, behavioural data and economic influence. A platform did not need to own the newspaper, television station or record label to exercise considerable power over distribution. Controlling the gateway through which people discovered content could be more valuable than producing the content itself.

The same pattern appeared across the digital economy. Cloud platforms became gateways to computing infrastructure, app stores became gateways to mobile software and online marketplaces positioned themselves between businesses and customers.

Artificial intelligence initially appeared to fit comfortably within this tradition. Large models absorbed enormous quantities of human-produced information and created a new interface through which people could retrieve, transform and generate knowledge.

The technology is now crossing a more consequential boundary because it is beginning to position itself between human intention and the execution of that intention.

AI Is Moving From Information Towards Action

A chatbot waits for a question and produces an answer. An agent receives an objective and attempts to pursue it. Although the distinction may sound modest, its economic consequences are enormous.

OpenClaw is best understood not as a new foundation model but as an operating architecture for AI. It connects models to persistent memory, communication channels, tools and machines. Its official description promotes an assistant capable of managing email, organising calendars and checking users in for flights through familiar messaging platforms.

These systems can remain active, preserve information between interactions and operate with access to files, credentials and external services. Researchers have consequently argued that Claw-like agents should increasingly be treated as computer systems rather than merely conversational products. They can install software, maintain state, schedule tasks and mediate interactions between other systems.

Astra represents progress at a different layer. Rather than creating the surrounding agent architecture, OpenAI has trained the model itself to perform computer-based work more effectively. The company says Astra can complete forms, update customer records, organise calendars, conduct research, build websites, operate professional software and troubleshoot problems appearing on screen.

It’s important to note, neither OpenClaw nor Astra invented computer control. What is changing is the combination of capability, persistence, access and reliability. OpenClaw showed what an always-available agent could look like, while Astra seeks to provide more of the intelligence required to make such systems dependable.

The destination is no longer an AI that merely explains how to perform a task. It is an AI to which the task itself can increasingly be delegated.

Control of Information Is Becoming Control of Execution

This transition could produce a profound shift in platform power.

Google’s historic position came partly from controlling the gateway through which people found information. Social-media platforms became influential because their algorithms determined which information people encountered. The emerging AI agent could occupy an even more consequential position because it may determine how a task is performed.

If someone asks an agent to organise a holiday, the system could decide which websites to search, which hotels to compare, which reviews deserve attention and which options should be excluded. It might eventually recommend and execute the transaction.

If the same system is asked to administer part of a business, it could choose software tools, suppliers, information sources and working methods. It would no longer be merely informing a decision. It would be shaping the pathway through which the decision becomes an action.

The emerging agent economy is competing to become the infrastructure through which those decisions are executed. This represents a deeper form of mediation because the platform is moving closer to the point at which intention becomes reality.

From Capturing What Humans Produce to Capturing What Humans Do

A separate development reveals the same transition from another direction.

A Guardian investigation published in June examined factory workers in India who were asked to wear head-mounted cameras while performing manufacturing tasks. These first-person recordings can provide valuable training material for AI and robotics systems attempting to learn how humans manipulate physical objects and complete complicated sequences of work. The investigation also raised questions about consent, surveillance, worker understanding and the distribution of the economic value created from those recordings.

The first generation of generative AI learned extensively from things humanity had already produced, including books, websites, photographs, software, music and conversations. The emerging generation increasingly needs to understand how humans act.

For robotics, this can involve observing how hands manipulate physical objects and how workers respond when something unexpected happens. For computer-use systems, it can involve learning how people navigate software, interpret interfaces and recover from errors.

The valuable dataset is therefore expanding from human output to human behaviour. AI companies are no longer interested solely in what people have created. They increasingly need to capture the process through which people create, decide and act.

Human Workflow Is Becoming an Asset

Every profession contains substantial amounts of tacit knowledge that never appear in a formal job description.

An experienced administrator knows which system usually fails and how to work around it. A developer learns which error messages matter and which can safely be ignored. A factory worker may recognise from touch or sound that a component has been assembled incorrectly. A teacher notices subtle signs of misunderstanding before a student explicitly asks for help.

Traditionally, much of this knowledge disappeared into the completion of the task. AI creates the possibility of observing, recording and converting parts of that process into reusable data.

Once a workflow can be captured, it may become training material. Once enough examples have been collected, parts of that workflow may become reproducible by machines.

This creates an uncomfortable economic possibility because a worker can produce value twice. The worker first performs the task and then, through the recording of that performance, may help train a system capable of reproducing parts of it.

Questions of ownership consequently become more complicated. If somebody’s expertise becomes valuable training material, it is reasonable to ask whether that knowledge should be treated merely as another by-product of employment. It is equally necessary to ask whether workers understand how their recordings will be used, whether they can refuse and whether they should retain an economic interest in the systems their behaviour helps to create.

These questions are not peripheral to the development of AI. They concern how the benefits of automation will be distributed and whose labour will subsidise the next generation of machines.

The Pursuit of Control Has Not Disappeared

There is no single movement with one agenda. Researchers, governments, open-source developers, start-ups, workers and multinational technology companies have different interests.

AI agents can also increase human agency. They may remove repetitive administrative work, make sophisticated capabilities accessible to people without specialist skills and allow small organisations to accomplish tasks that once required much larger teams.

However, the economics of previous digital platforms should make us cautious about where control accumulates.

The scarce resource of the social-media era was attention. The scarce resource of the cloud era was computing infrastructure. The scarce resource of the early generative-AI era was machine intelligence and the computing power required to deliver it. The scarce resource of the agentic era may increasingly be permission. An effective agent needs authority to access files, read emails, use browsers, install software, communicate with organisations and potentially spend money. These permissions transform intelligence into practical power.

A model without access may be impressive but largely contained. A less intelligent model with memory, credentials, browser control and permission to communicate with external systems may be considerably more powerful.

The important equation is therefore no longer simply that intelligence produces capability. In the agentic era, practical power is closer to intelligence multiplied by access, persistence and autonomy.

The Most Powerful AI May Not Be the Smartest Model

This also means that benchmark results may eventually tell us less about the true distribution of AI power than we imagine.

The article about Astra acknowledges that its independent general-intelligence score increased only marginally from 60.9 for GPT-5.6 Sol to 61.2. Its larger gains appear in computer use, automation, terminals and specialist workflows.

Astra may not therefore represent an equally dramatic advance across every form of intelligence. Its importance lies in converting existing intelligence into more reliable action.

This is why the convergence between OpenClaw-style systems and computer-native models deserves attention. OpenClaw demonstrated the persistent operating architecture. Astra represents a stronger model capable of working inside such an architecture.

The inflection point will arrive when these layers become reliable enough for ordinary users to stop supervising every individual action. At that stage, the interface to computing begins to change. Rather than operating applications ourselves, we express an intention and permit an intermediary to operate them for us.

This Creates a Security Problem We Have Barely Started to Solve

Granting an AI agent greater agency also increases the consequences of failure.

A hallucinated sentence can be corrected or ignored. A mistaken action can submit a form, expose confidential information, delete a file, alter hundreds of customer records or send a message to the wrong person.

Claw-like agents create a particularly difficult security problem because they may operate continuously while retaining access to credentials, files, communication channels and external services. Researchers describe this as a substantially enlarged attack surface involving malicious skills, compromised plugins, persistent-memory attacks and the possibility of failures cascading between agents.

The underlying model is only one part of the problem. A well-behaved model operating within a badly designed permission system can still create substantial risk. Likewise, a trusted agent may become dangerous if one of its plugins, retrieved documents or external instructions has been manipulated.

Traditional AI-safety questions about harmful content, bias and hallucination are therefore no longer sufficient. Agentic systems require the protections developed through decades of operating-system and cybersecurity engineering, including isolation, least privilege, authentication, access control, audit trails, revocation and sandboxing.

The crucial safety question is shifting from what an AI is permitted to say towards what it is permitted to do.

Astra Makes the Question More Urgent

Astra illustrates how quickly the capability frontier is moving. OpenAI describes it as its first broadly deployed model to reach the Critical cybersecurity capability level under the company’s Preparedness Framework.

According to OpenAI, Astra can identify previously unknown vulnerabilities and develop methods of exploiting sophisticated systems when it has the necessary tools and access. During testing, the model reportedly discovered and used two previously unknown software vulnerabilities, which OpenAI says are being disclosed to their maintainers.

The company simultaneously reports substantial improvements in alignment and computer-use safety. Astra performed better than GPT-5.6 Sol on internal safety evaluations and was less likely to exceed the authority provided by the user. However, OpenAI also acknowledges that the model’s written reasoning can be harder to monitor and that additional safeguards may interrupt legitimate activity.

Both sides of this development matter. The model is becoming more capable of acting, while its developer is attempting to make that increased capability safer. This tension between capability and control will define much of the next phase of AI.

The Battle May Ultimately Be Over the Agent Layer

The competition between foundation models receives enormous attention, but the more strategically important contest may ultimately concern who controls the agent surrounding those models.

Foundation models could become increasingly interchangeable. An agent might use one model for coding, another for research and a cheaper model for routine administration. The lasting commercial advantage could therefore migrate towards the system holding the user’s identity, memory, preferences, relationships, credentials, permissions and behavioural history.

That system would know more than what the user asked yesterday. It could learn how the person operates, which decisions they normally make and which organisations they trust.

This could become one of the most valuable positions in the technology industry. The company controlling the agent layer could mediate relationships between individuals and banks, retailers, employers, governments, healthcare providers and software companies.

The great platform battle of the next decade may not simply be about who develops the most intelligent model. It may be about who becomes trusted to act on our behalf.

Or Is This the Same Pursuit of Control?

There is also considerable continuity. Every major generation of digital technology has created a new intermediary and produced a struggle over who controls it.

Agentic AI may organise access to action itself. The principal danger is not necessarily a science-fiction scenario in which a single artificial intelligence suddenly seizes control. A more plausible concern is the gradual concentration of authority.

People may willingly delegate small decisions because doing so is convenient. Those decisions accumulate into workflows, the workflows develop into dependencies and the dependencies eventually allow a small number of systems to mediate a significant proportion of digital economic activity.

Control does not always arrive through force. Sometimes it arrives because delegation is easier.

The AI debate has spent years asking whether machines will become more intelligent than humans. That is no longer the only question that matters.

We must also ask who owns the infrastructure through which that intelligence acts, who controls its permissions, whose interests influence its decisions, who owns the behavioural data used to train it and whether individuals can genuinely withdraw the authority they have delegated.

As we wrap up this deep dive, the first generation of generative AI learned extensively from what humanity produced. The emerging generation is learning how humanity works. The first generation answered our questions. The next generation will increasingly perform our tasks.

The defining struggle of the agentic era may therefore concern more than intelligence. It may concern who is granted permission to turn our intentions into actions and who ultimately controls the infrastructure standing between the two.