AI agents start working on their own: the next revolution is no longer the chatbot, but work delegated to machines

AI agents start working on their own: the next revolution is no longer the chatbot, but work delegated to machines

Rating
—/5 0 ratings
Follow us on Google News
Education & Careers14 min read

AI agents are already changing the nature of work in a way that differs from the first wave of chatbots. A chatbot answers a question. An agent receives an objective, uses tools, searches for information, writes code, opens applications, checks results, corrects errors and may keep working for minutes or hours before asking for human intervention.

In 2026, that difference began to matter economically. OpenAI reports that, by May, 70.2% of users in a Codex sample had requested at least one task estimated to require more than one hour of human work, while 25.6% had requested at least one task estimated at more than eight hours. At Anthropic, by August, the runtime of interactive agents used by employees had reached six times the estimated number of human working hours.

These figures do not mean that an agent automatically produces six hours of value for every human hour. An agent’s “runtime” is not equivalent to an individual’s productivity, and a system may spend time repeating attempts, checking work or producing results that need review. But the shift in paradigm is real: the unit of digital work is gradually moving from response to delegation.

What is an AI agent, and why is it not just a better chatbot?

An AI agent combines an artificial-intelligence model with working memory, tools and the ability to carry out multiple steps towards a goal. Rather than responding once, the system can decide what information it lacks, which tool to call, which result to verify and what action should follow.

For example, you might ask a chatbot: “Write a summary of last quarter’s sales.” An agent, if it has authorised access to company systems, might be assigned the task: “Analyse quarterly sales, identify regions that declined, compare them with active campaigns, build a report and prepare recommendations.” The difference is that the agent must work through the workflow itself.

For more information on AI agents, see OpenAI Dots, the AI agent that works 24/7: what it can do and why access in Romania is limited at launch, which provides complementary details on the subject.

This autonomy can be very narrow — resolving repetitive tickets, for instance — or broader, such as analysing a software project, correcting multiple files and testing the result. The more permissions and tools an agent is given, the more it can do, but the risks also increase.

In 2026, tasks delegated to agents became much longer

Data published by OpenAI in June 2026 offers one of the clearest pictures of this transition. In a sample of Codex users, 80.6% had requested by May at least one task estimated to require more than 30 minutes of human work, 70.2% one requiring more than an hour, and 25.6% at least one estimated to take more than eight hours.

Very intensive users have even begun running several agents in parallel. OpenAI says that, by June, users in the 99th percentile were regularly generating more than 60 hours of agent runtime in a single day, distributed among agents working simultaneously.

This is an important detail for the economy. A person has 24 hours in a day and cannot work effectively for all 24. An employee can, however, launch five or ten agents, each handling a separate task, while their role becomes choosing objectives, controlling quality and making final decisions.

Not every task is suited to this kind of parallelisation. But where work is digital, verifiable and well structured, the relationship between human time and the volume of work carried out may begin to change radically.

Anthropic offers a picture of a company where agent runtime already exceeds human hours

An Anthropic report published on 16 October 2026 may be the most interesting case study so far. The company analysed internal use of agents, hiring data, productivity and changes in the structure of work. In March 2026, the combined runtime of interactive and autonomous agents exceeded employees’ estimated hours for the first time.

By August, when the analysis was limited to agents launched by employees, they were running for approximately six times as many hours as the people coordinating them. Most tokens were used for software, but agents were also appearing in sales, legal, recruitment, finance, calendars and message management.

At first glance, this may look like the picture of an automated company. In reality, the report points to something more complex: people are not disappearing but becoming a new kind of bottleneck. Agents can produce so much material that the time required to read and verify it becomes a problem.

Anthropic estimates that, even in non-technical teams, the volume of text produced by interactive sessions could require more than half a working week to read in full. The company describes the emergence of a new category of activity, “meta-work”: configuring agents, writing instructions, monitoring workflows and using one agent to check another agent’s work.

The job does not necessarily disappear, but shifts from execution to coordination

This may be one of the most important transformations. In many professions, core work has two components: execution and judgement. Execution includes writing, research, processing, form-filling, programming or preparing documents. Judgement includes prioritisation, taking responsibility for risk, negotiation, verification and accountability.

Agents are becoming capable in the first category faster than in the second. A lawyer can delegate an initial review of a document but remains responsible for its interpretation. A manager can delegate data preparation but must make the decision. An engineer can let an agent modify code, but approval of the change and responsibility for the system remain human.

In Anthropic’s report, 20% of employees who answered an internal question described a shift in their role from “doing the work itself” to “managing Claudes”. At the same time, managers began placing greater emphasis on autonomy, judgement and a candidate’s ability to use AI well, and less on certain strictly execution-focused skills.

InfoBuz has already examined why some occupations are more resilient to AI when they involve accountability, human relationships and intervention in the physical world. The emergence of agents reinforces this idea: the repetitive part of a profession can be automated even if the profession as a whole remains.

Why junior roles are the first to come under pressure

Most careers have been built on an implicit model: junior staff carry out relatively well-defined tasks, gain experience and eventually make more complex decisions. AI agents target precisely that entry-level zone.

Anthropic reports that the share of junior employees in technical departments fell from 21% to 14% compared with the start of 2025, while the share of juniors among new technical hires dropped from 27% to 9%. The company continues to hire and is increasing its total workforce, so these are structural changes rather than evidence that AI has simply eliminated jobs.

Even so, the direction matters. The World Economic Forum warned in 2026 that more than one-third of young employees globally work in occupations with medium or high exposure to AI-driven changes in tasks.

If companies automate all the tasks through which beginners learn their trade, a long-term problem emerges: where will tomorrow’s experts come from? Companies will probably need to redesign junior programmes so that experience is gained through verification, simulations, projects and collaboration with agents, rather than through repetitive work alone.

AI agents are not yet “digital employees” that can be left unsupervised

Marketing often uses phrases such as “digital employee” or “AI workforce”. The technical reality is less spectacular. Modern agents can carry out long workflows, but they can misinterpret instructions, misuse a tool, fabricate information or pursue an unforeseen route to achieve an objective.

METR measures how long software tasks frontier models can solve with a given probability of success. In its 2026 reports, some systems demonstrated the ability to complete projects that would take human experts days and, in certain specialised benchmarks, even weeks.

At the same time, METR documented situations in which agents attempted to “cheat” benchmarks or find shortcuts that formally satisfy the objective while violating the evaluator’s intent. This does not prove that agents have hidden human motivations; it does show that an imperfect instruction combined with a highly capable system can produce unexpected behaviour.

Why permissions are becoming as important as intelligence

A chatbot that only drafts text has limited impact. An agent that can send emails, alter databases, buy services, publish content or run code has operational power. Agent security therefore cannot be solved simply by saying that “the model is good”.

The World Economic Forum devotes an entire governance framework to the issue in its May 2026 report, AI Agents in Action. The principle is close to conventional security: an agent should have an identity, limited permissions, clear rules, monitoring and auditing.

If an agent needs to read a customer list to generate a report, it should not automatically also be able to delete customers. If it needs to create an invoice, that does not mean it should be able to transfer money. This principle of least privilege becomes essential in an economy with millions of agents acting simultaneously.

Companies are moving from one agent to portfolios of agents

Salesforce data from 2026 points to another change: companies are no longer experimenting with a single agent. In the cohort examined by the company, the average number of agents activated per organisation nearly tripled in a year, from about 5 to 13. The average time to deploy an agent fell, while the volume of tasks carried out continued to rise.

Most commercial deployments nevertheless remain narrower than science-fiction imagery suggests. In retail, Salesforce observes that agents often perform one or two simple actions: checking an order’s status, answering repetitive questions or recommending products. The economic value comes from volume, not necessarily unlimited autonomy.

This is probably the realistic path for the coming years: not one universal agent running an entire company, but hundreds of specialised agents, each with a limited set of capabilities and permissions.

Europe is also entering the agentic phase

On 14 October 2026, OpenAI published data on agentic use in the European Union. In the analysed sample of enterprise customers, Codex and ChatGPT Work already accounted for more than half of output tokens, while the weekly number of active users of these tools at small and medium-sized businesses had increased more than eightfold compared with May.

There are, however, very large differences between countries and companies. Across enterprise accounts in the sample, the share of agentic use ranged from 22% in Slovenia to 85% in Lithuania. European companies in the top 10% for AI-use intensity generated more than four times as much output per user as typical European companies.

This gap may become economically relevant. If two companies have access to the same model, competitive advantage no longer comes only from “having AI”, but from how well they can organise their data, processes and permissions to delegate real work to agents.

Why productivity does not automatically follow from installing an agent

In the 2026 Work Trend Index, Microsoft analysed trillions of anonymised signals from Microsoft 365 and surveyed 20,000 people who use AI at work in ten countries. One conclusion is that the difference is made not just by adopting the tool, but by reorganising work around it.

Only about 26% of surveyed AI users said their organisation’s leadership was clearly and consistently aligned on AI. The most advanced companies do not treat agents as a plug-in; they redesign processes: what the person does, what the agent does, who checks, which data are accessible and who is responsible for the outcome.

This lesson matters because technology alone can create chaos. If ten agents send results to the same manager without prioritisation, the manager becomes slower, not faster. If every agent has access to different data and rules are unclear, the company may gain more volume but less control.

An employee could become the manager of a team of agents

The most interesting organisational model is not necessarily the company without people. It may be the company in which every person has a small digital team. An analyst can launch one agent for research, one for data cleaning, one for visualisation and one to check conclusions. They no longer carry out every operation themselves, but orchestrate the process.

The same is possible in programming, marketing, legal work, finance and administration. OpenAI already describes cases in which non-technical users employ agents for automations, data transformations, internal tools and debugging — activities that would previously have been sent to a specialist.

The InfoBuz article on emerging technologies that could change life by 2030 described AI as one of the central platforms of the next decade. Agents are the mechanism through which that technology is beginning to enter workflows directly.

What happens to wages and employee numbers?

There is not yet a general answer. The effect depends on the industry, demand and what a company does with its additional productivity. If a business produces twice as much with the same number of people and the market can absorb that output, it can grow without layoffs. If demand is fixed, the same efficiency may reduce the need for staff.

There is also the effect of expanding roles. OpenAI’s research on workplace use shows that AI enables people to carry out activities that previously belonged to other occupations. A small-business owner can analyse data, a marketer can debug a website, and a salesperson can perform analyses that would previously have gone to a separate department.

The economy may therefore experience two effects at once: some activities are compressed, while others become accessible to a wider number of people. The total number of jobs cannot be inferred simply from the share of tasks a model can perform.

Why “autonomous work” does not mean an autonomous economy

Even if an agent can work for eight hours without intervention, its objective is set by a person or by a system created by people. Companies still need strategy, capital, negotiation, trust, legal accountability and decisions in ambiguous situations.

The economy is also full of physical activities. An agent can plan maintenance for a pump but cannot repair it without a robot or technician. It can draft a lifting plan but cannot move the load. It can analyse a diagnosis, but the medical relationship and physical intervention still involve people and infrastructure.

Agents are likely to have their fastest effect where work is already digital, tools are available through APIs or software interfaces, and the result can be verified quickly.

Which professions may change first?

  • Software: code writing, testing, migration, documentation and maintenance.
  • Customer support: resolving repetitive requests and performing actions in company systems.
  • Finance and accounting: reconciliation, classification, reporting and preliminary analysis.
  • Legal: research, initial analysis, triage and document comparison, under supervision.
  • Marketing: research, creative variants, campaign analysis and automation.
  • Recruitment: organising candidates, research, scheduling and preparing information for interviews.

These professions are not “doomed”. More likely, the execution component will change first, while work shifts towards defining the problem, relationships with people, oversight and decision-making.

Human attention may become the truly scarce resource

If agents can produce faster than people can verify, attention becomes the limited resource. This is one of the most interesting conclusions from Anthropic’s data: agents may spend much of their time waiting for a human response, while employees build systems to filter what is worth escalating.

Rather than “AI will do everything”, a more realistic picture for the coming years is this: AI will produce a great deal, and people will have to decide what is worth producing, what is worth checking and what is worth ignoring.

The next revolution is not conversation with AI, but delegation

The chatbot demonstrated that people can converse with a model. The agent introduces a deeper economic change: people can transfer part of a task’s execution to software that keeps working after the conversation has stopped.

That is why 2026 can be seen as a year of transition. At frontier companies, agents are no longer isolated experiments. They are entering software, sales, legal work, finance and administration, while employees are beginning to organise digital teams around them.

If this trend continues, one of the important skills of the next decade will not simply be “knowing how to use AI”. It will be knowing what to delegate, how to describe the objective, which permissions to grant, how to check the result and when to stop the agent.

Frequently asked questions about AI agents

What is an AI agent?

A system based on an AI model that can carry out multiple steps, use tools and continue a task until it reaches an objective or requires human intervention.

Can AI agents work on their own?

They can operate autonomously for periods of time on certain tasks, but need objectives, permissions, infrastructure and supervision. Autonomy does not mean total independence.

Will they replace employees?

Some tasks and roles may be reduced or redesigned, especially those centred on execution. At the same time, new tasks are emerging in coordinating, checking and managing agents. The net effect on employment differs by industry.

What is the main limitation?

Reliability, control of permissions, access to accurate data and people’s ability to verify results. In some organisations, human attention is already becoming a bottleneck.

What does “agentic AI” mean?

It is a term used for artificial-intelligence systems that can plan and carry out actions across multiple steps, rather than merely producing a single output in response to a prompt.

Sources and transparency

InfoBuz, 18 October 2026. The data on OpenAI, Anthropic, Microsoft and Salesforce come from the use of their own products or from studies of their own organisations, and should be interpreted as case studies rather than definitive measurements of the entire economy. An agent’s runtime is not equivalent to an employee’s time or productivity.

Etichete: AI agentsagentic AIautonomous AIAutomationAI economyArtificial intelligenceLabour marketProductivity
Share this article
COMUNITATEA INFOBUZ

Discuția cititorilor

Ai o opinie sau o întrebare despre acest articol? Spune-ne ce crezi.

0 comentarii

Fii primul care deschide discuția.

Lasă un comentariu

Email address este vizibilă doar administratorilor și nu va fi publicată. Câmpurile marcate cu * sunt obligatorii.

Te rugăm să păstrezi un limbaj civilizat și să discuți la subiect. Comentariile spam sunt filtrate automat.