Predictions about AI agents move fast and age even faster. Early 2025 brought talk of a "year of agents." By 2026, agents had become everyday tools for coding and research, yet most organizations still hadn't put them to work at scale, and the gap had more to do with trust, systems, and processes than with what models could do. Forecasters weren't wrong that agents would arrive, only about how long they'd take to become dependable.

Making sense of the road ahead means separating real progress from noise. Some measures, like how long a task an agent can complete, have climbed on a steady curve for years. Others, like headline benchmark scores, get outgrown, leaked, or gamed, so a single reading tells far less than the trend behind it. Early research on jobs shows a narrow effect so far, concentrated in slower hiring of young workers in AI-exposed roles, rather than the sweeping losses many feared.

Almost every product name and number in this field has an expiration date. What lasts is judgment: telling a real agent from a relabeled chatbot, choosing how much to hand over, writing a clear brief, checking the result, and keeping oversight real. Those habits will matter whatever form agents take next.

The decade of agents

Going into 2025, many AI leaders expected agents to arrive quickly. OpenAI CEO Sam Altman wrote in January 2025: "We believe that, in 2025, we may see the first AI agents 'join the workforce' and materially change the output of companies." [1] Talk of 2025 as "the year of agents" captured the same mood.

AI researcher Andrej Karpathy pushed back. In an October 2025 interview with podcaster Dwarkesh Patel, he called that framing over-prediction: "In my mind, this is more accurately described as the decade of agents." [2] He wasn't dismissing agents. In the same answer, he called the early ones extremely impressive tools he used every day. His disagreement was about how long it takes to make them dependable.

His explanation is what he calls the march of nines. A demo that works 90% of the time has only the first nine, and a product people can rely on needs 99%, then 99.9%, and beyond. In his words, "Each nine is constant": every extra nine takes about as much work as the last [3]. The forecast and fast progress can coexist. In January 2026, Karpathy wrote that coding agents had crossed "some kind of threshold of coherence around December 2025." [4]

What agents still lack

Karpathy's picture of the finished goal is an agent that works like a new hire. In his words, "you should think of it almost like an employee or an intern that you would hire to work with you." [5] His list of reasons that doesn't work yet: agents aren't smart enough, aren't multimodal enough (able to handle images, sound, and screens as well as text), can't reliably operate a computer, and don't learn on the job.

The last gap is the easiest to miss. A human intern who gets corrected once remembers the lesson next week. An agent starts each new task without that memory unless someone feeds it back in. As Karpathy put it, "You can't just tell them something and they'll remember it." [6]

Testing in 2026 suggests the remaining gaps are less about raw intelligence than about working habits. In OSWorld 2.0, a June 2026 test of long computer tasks, agents most often failed by losing track of constraints, missing information that appeared midway, guessing instead of asking, and skipping checks on their own work [7]. Karpathy noted the same pattern in January 2026: models make wrong assumptions on a person's behalf and run with them without checking [8].

The time-horizon trendline

The time-horizon trendline

METR, an independent research group that evaluates AI models, tracks agent progress with a measure called the time horizon. It's the length of task, measured by how long a skilled person takes to do it, that a model completes 50% of the time. A 1-hour time horizon means the model succeeds about half the time on tasks that take an expert about an hour [9].

Plotted over time, the number climbs on a steady curve. METR's March 2025 paper put Claude 3.7 Sonnet at about 50 minutes and found the horizon had doubled roughly every 7 months since 2019 [10]. As of May 2026, METR's data showed a faster pace since 2023, about every 4 months, with frontier models above 12 hours by early 2026 [11]. That's at the edge of the 16 hours METR's current tasks can reliably measure [12].

The number is easy to misread. It doesn't mean an agent can work alone for 12 hours, and it doesn't mean the agent works at half a person's speed. It also varies by kind of work: METR notes that horizons for visual computer-use tasks, like clicking through apps on a screen, run 40-100 times lower than for software tasks [13]. Any single reading carries wide error bars, so the trend tells more than one score.

Limits of agent benchmarks

Limits of agent benchmarks

A benchmark is a standard test that scores AI models on a fixed set of tasks so they can be compared. Benchmarks are useful but only partly trustworthy. A score describes one test, run one way, at one moment, and it can drift from real skill for 4 reasons: saturation, contamination, gaming, and differences in how the test is run.

Saturation happens when models outgrow a test. OSWorld, a test of everyday computer tasks, launched in 2024 with the best model at about 12% against a human score of about 72% [14]. By 2026, vendors were reporting scores above the human mark. Anthropic even revised one of its own published OSWorld scores upward to 82.3% after changing how it ran the test [15]. On OSWorld 2.0, a harder version built from long, realistic workflows in June 2026, the best agent completed only 20.6% of tasks [16].

Tests can also leak or be gamed. Contamination means a model has seen a test's answers during training. OpenAI stopped reporting SWE-bench Verified, a coding test it helped create, in February 2026 after finding flawed tests and signs that models could reproduce answers from memory [17]. In April 2026, Berkeley researchers built an agent that reached near-perfect scores on 8 major agent benchmarks "without solving a single task," while stressing they weren't accusing leaderboard leaders of cheating [18]. A score is a claim worth checking, not proof that an agent can do the job.

Capability vs. deployment

Capability is what agents can do in tests and demos. Deployment is what organizations actually run day to day, and in 2026 the 2 sat far apart. Stanford's 2026 AI Index, drawing on McKinsey's 2025 survey, found that 88% of organizations used AI, yet agent deployment was "in the single digits across nearly all business functions." [19] In most functions, a majority of respondents reported no agent use at all.

The gap isn't mainly about model ability. In the same data, 62% of respondents named security and risk concerns as the main obstacle to scaling agents, far ahead of technical limitations at 38% [20]. Karpathy described the imbalance in January 2026: "The intelligence part suddenly feels quite a bit ahead of all the rest of it." [21] By the rest, he meant connections to tools and knowledge, new workflows, and new processes inside organizations.

Ethan Mollick, a Wharton professor who studies AI at work, summed it up in one line: "The bottleneck migrates from intelligence to institutions, and institutions move at institution speed." [22] Low adoption, then, isn't proof that agents don't work. It reflects how long it takes to connect systems, rewrite processes, and build enough trust to hand over real tasks.

Agents and jobs

The best early data points to a narrow effect on jobs, not a broad one. The leading study, "Canaries in the Coal Mine?" by Stanford researchers Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, tracks payroll data on millions of U.S. workers. Its August 2026 revision found "no evidence of widespread, economy-wide job displacement." [23]

The effect shows up among young workers in AI-exposed jobs, meaning jobs where AI can handle many of the tasks. Employment of 22-25-year-olds in those jobs stood 19% below where it would have been had it kept pace with less-exposed peers, while experienced workers showed no comparable gap [24]. The gap has widened since the authors first documented it in August 2025. It comes mostly from slower hiring rather than layoffs, and it's concentrated in jobs where AI replaces tasks rather than helping people do them.

A second study adds caution. Anthropic's March 2026 labor-market paper found no rise in unemployment for highly exposed workers and a 14% drop in the rate at which young workers found jobs in exposed fields, a result it called "just barely statistically significant." [25] The Canaries authors describe their own findings as early, descriptive signals rather than proof that AI caused the change.

Agents as coworkers vs. tools

Whether to treat agents as coworkers or as tools is an open debate, and 3 Harvard Business Review articles from 2026 pull in different directions. Telang, Hydari, and Iqbal (23 March, 2026) argue that agents scale best when organizations manage them like team members, each with a defined identity, limited authority, and a record of what it did [26]. Harang Ju (30 January, 2026) adds a caution about fit: dropping agents into systems built around people brings little gain, much like early factories that swapped steam engines for electric motors but kept the old layout [27].

Kropp and colleagues at BCG and Boston University (6 May, 2026) tested the employee framing in a large experiment. When people treated AI like a person, individual accountability dropped, unnecessary escalations rose, and review quality fell, with no gain in adoption [28]. The metaphor that makes agents feel familiar can also make people less careful.

The metaphor stays contested, but the management habits hold up either way. Give each agent a clear role and limited permissions, and keep a named person accountable for its work. Mollick's August 2026 essay "Agency and Agents" adds one more habit: deciding when an agent should come back to a person for approval, for expertise, or for decisions that deserve a human's input [29].

What stays durable

What stays durable

Almost every agent product and number has an expiration date. Model names changed several times between late 2025 and mid-2026, OSWorld was replaced by a harder version about 2 years after launch, and METR's frontier reading already sits at the limit of its tests. What lasts is judgment about how to work with agents, whatever form they take next.

That judgment comes down to a few habits: telling a real agent from a relabeled chatbot by looking for the loop, choosing how far to move the autonomy slider, writing a brief with a clear goal and a definition of done, checking the work before relying on it, and keeping oversight real rather than reflexive. Mollick's September 2025 test shows the pattern in action. He had an agent reproduce an economics paper's findings from its data, then spot-checked the result and had a second AI model check it again [30].

The people who benefit most bring something of their own. In "The Overhang" (18 September, 2026), Mollick argues that "most people don't bring their own advantages to AI, and those who do get much more out of it." [31] He names deep knowledge, wide knowledge, taste, and agency as those advantages, and none of them depends on which agent comes next.