What can AI agents do, and can they run a company?

AI agents in late 2026 can work for hours, coordinate in teams and complete a growing share of well-defined digital tasks. They still fail to finish many long workflows, and they cannot hold legal or commercial accountability. They can do much of the work inside a company, but people still have to decide and answer for it.
This article is dated on purpose. Agent capability changes fast enough that a claim made in spring can be wrong by autumn. Everything below describes the evidence as it stands on 16 November 2026, with the source and date given for each figure.
What can AI agents do today?
An AI agent is a model that is given a goal, a set of tools and room to act over many steps. It plans, uses software, checks its own output and continues until it stops or is stopped. For how groups of them are organised, see what is an agent swarm.
Four things changed in 2026.
Agents work for longer. A summary by NYU Shanghai, dated 17 April 2026, of Anthropic’s 2026 Agentic Coding Trends Report describes tasks that run for hours to days. The summary is a secondary source for the report.
Agents work together. The same summary describes single agents becoming coordinated teams. Y Combinator’s Fall 2026 Requests for Startups, archived 1 September 2026, includes a request called Multiplayer AI, for teams of people who share live agent sessions as agents run for hours, days and weeks.
Agents complete more real work. The Remote Labor Index, a project of the Center for AI Safety and Scale AI, showed 41.25% full automation of real remote-work projects on its site when read on 16 November 2026, with a doubling pace of about 3.3 months. The original result from October 2025 is reported as 2.5%. The figure is the project’s own reading on that date, and its current methodology is outside this article.
Agents are gaining the means to transact. Stripe’s documentation describes the Agentic Commerce Protocol as an open standard, created by Stripe, OpenAI and Meta, for agents completing purchases on a buyer’s behalf.
How reliable are AI agents at finishing a job?
Less reliable than the capability list suggests. The clearest numbers come from the research section of the State of AI Report 2026, published by Air Street Capital in October 2026.
On OSWorld 2.0, a test of computer workflows, the best model earned 77.7% partial credit and fully completed 44.3% of the workflows. On FrontierChallenge, the best result was 87.9 out of 100 on requirements, with 20.6% of tasks fully completed.
Each pair has a high score for partial work and a much lower one for finished work. An agent can satisfy most of what a task asks and still not deliver the task. In a business, a job that is nearly done is often worth nothing until someone finds the missing part and fixes it.
Anthropic has shown the same pattern in a small real setting. In Project Vend phase two, an AI shopkeeper improved markedly with newer models. Anthropic’s own conclusion is that the gap between capable and completely robust remains wide.
How long can an AI agent work on its own?
Long enough that length is no longer the main limit, and not yet with dependable results at the long end.
METR tracks what it calls the 50% time horizon: the length of task an agent can complete at half reliability. Its time-horizons page, last updated 8 September 2026, flags measurements above 16 hours as unreliable with the current task suite.
That definition carries two cautions. Half reliability means the agent fails as often as it succeeds at that length, which no business would accept for important work without checks. And the measuring tools themselves are running out of range, so claims about very long autonomous runs are hard to verify.
An agent that can run for a day is useful. An agent that can be trusted with a day’s work unsupervised is a stronger claim, and the evidence for it is thinner.
What can AI agents not do yet?
Finish reliably without checks. The completion figures above describe this directly.
Exercise judgment where the rules run out. The State of AI Report 2026 says agents can do research engineering but score poorly on research judgment. Deciding which question is worth asking is a different skill from carrying out the work once it is asked.
Be accountable. Gartner said on 2 October 2026 that most agents are not reliable enough for autonomous use in business-critical settings, and that human accountability cannot be removed. It predicts that 40% of agentic AI projects will be cancelled by the end of 2027.
Stand in for a legal person. Bloomberg Law reported on 2 April 2026 that no jurisdiction lets a company exist without a traceable natural person. Whatever an agent does, a human being is answerable for it.
Supply access. An agent can write a thousand good emails. It cannot be the person a buyer already trusts, and in many markets that trust decides who gets a meeting.
Why do the numbers on AI agent reliability disagree?
The figures 41.25%, 44.3%, 20.6% and 77.7% all describe agent performance in the same year. They measure different things.
Some scores award partial credit and others count only fully completed tasks. Some test computer use, some test freelance-style projects, some test software engineering. Some are run by independent evaluators and some are published by the project that designed the test.
Before relying on any agent benchmark, check three things:
- whether the number counts finished work or partial work;
- who ran the test;
- how close the tasks are to the job you have in mind.
All of them point the same way. Capability is rising quickly, and full completion trails behind partial credit.
Can AI agents run a company?
They can do a large part of a company’s work. Running a company also includes deciding what the company should do, committing money, making promises to customers and being answerable when something goes wrong.
On the evidence above, those remain human functions. Agents complete under half of long computer workflows in the State of AI figures. Evaluators say accountability cannot be removed. The law requires a person.
There are products that promise more than this, and they are examined in can a company run itself. The closest thing to a consensus in late 2026 is narrower: agents can carry the volume of the work if people hold the decisions.
How should a business divide work between agents and people?
The NYU Shanghai summary of Anthropic’s report gives two figures side by side. Developers use AI in about 60% of their work, and can fully delegate only 0 to 20% of tasks. Most of the value comes from supervised work, where a person stays close to what the agent is doing.
The working rule that follows: give agents work that is high in volume, digital, and checkable against a clear definition of done. Keep with people the decisions that set direction, commit resources or bind the company to someone else.
The scarce input is human. The State of AI Report 2026 describes the gap between people who get value from AI and people who do not as a human skill issue and not a technology issue. TechCrunch reported on 30 July 2026, citing a Christian & Timbers study, that about 2,000 US engineers were judged able to deliver enterprise AI returns consistently.
The practical limit on agents in a business is therefore often the supply of people who know how to direct them, check them and take responsibility for the result. A fuller account of that way of working is in what is agentic company building.
Where Frontmark stands
Frontmark’s division of labour follows the evidence above. Agents do the volume: research, demand tests, product and their own tools. One accountable operator makes each decision that sets direction, commits resources or binds a company, and answers for it.
The same evidence is summarised on the agents page. The access agents cannot supply comes from the collective.
Common questions.
Can AI agents run a company?
Agents can carry out much of the daily work of a company, including research, coding, support and routine operations. As of November 2026 they still need supervision to finish long tasks reliably, and they cannot be legally accountable. A company can run largely on agents, but people still have to make the judgment calls and answer for the outcome.
What can AI agents do in 2026?
Agents can work on a task for hours, use computers and software tools, write and test code, carry out research, and coordinate with other agents in teams. They are strongest on well-defined digital work with a clear definition of done. Capability is rising quickly, which is why any list like this should carry a date.
How reliable are AI agents?
Less reliable than their headline scores suggest. The State of AI Report 2026 says the best model earned 77.7% partial credit on OSWorld 2.0 but fully completed 44.3% of computer workflows. Agents often get most of a job right and still leave it unfinished, which matters when the job is the product.
How long can an AI agent work on its own?
Reports in 2026 describe tasks that run for hours to days. METR tracks a measure called the 50% time horizon, which is task length at half reliability, and flags its own measurements above 16 hours as unreliable with the current task suite. Long duration and dependable completion are separate things.
What can AI agents not do?
Agents cannot be accountable for a decision, hold a legal duty or be the person a customer or regulator holds to account. They score poorly on research judgment, according to the State of AI Report 2026. They also cannot produce trust and access in a market, which come from people who are already known in it.
Why do so many AI agent projects fail?
Gartner predicted in October 2026 that 40% of agentic AI projects will be cancelled by the end of 2027, and said success depends on reliability more than autonomy. Projects tend to fail when an agent is given business-critical work without checks, without a clear definition of done and without a person who owns the result.