Blog · November 10, 2026
Where agentic AI stands
What the words mean, what the tests show, and where use is still uneven. Every number below comes from a named source.
What the words mean
In February 2026, the OECD published a paper on how people define these terms. The authors are Luis Aranda and Kasumi Sugimoto. The paper gives this shared reading.
This AI is a system that can perceive and act in its environment with some autonomy. It can use tools to reach a specific goal. It can adapt when the inputs change.
Agentic AI, in that same paper, generally means several coordinated AI systems. They can split tasks, collaborate, and pursue a complex goal over a longer time. They are meant for more open settings, and for less human supervision.
The OECD also says that many developers have started using this kind of AI, and that use does not mean the technology is mature. People still want stronger security, privacy, and accuracy.
The skill is uneven
Stanford's Institute for Human-Centered AI published takeaways from its 2026 AI Index on April 13, 2026. Shana Lynch wrote the summary. It says capability is not even.
On a test called Terminal-Bench, the success rate of this AI on real world tasks rose from 20% in 2025 to 77.3% at the time of that writeup.
AI handling cybersecurity issues solved problems 93% of the time. In 2024 that figure was 15%.
The same summary says AI still lags at other work. That includes planning across many steps, and financial analysis. It also includes learning from video, making coherent realistic video, telling time, and some expert exams.
Robots succeed at only 12% of real household tasks, such as folding clothing or washing dishes.
Companies are using them unevenly
McKinsey published its Global Survey on the state of AI on August 25, 2026. The survey ran from May 4 to June 8, 2026. It had 1,719 participants in 97 nations. The data are weighted by each nation's share of global GDP.
McKinsey says 40% of respondents at large organizations report scaling this kind of AI. Large means more than $1 billion in annual revenue. A year earlier that share was 27%. At smaller organizations the share stayed flat at 22%.
About two in ten respondents report scaling software coding AI across the organization. At larger enterprises that share is 31%.
Chatbots are still the tool most often scaled across a whole company. McKinsey puts that share at 47%.
Broader use has not shown up as broader profit. McKinsey says 37% of respondents attribute some earnings impact to AI, measured as EBIT. That share is about the same as the year before.
Nearly a third of respondents, 32%, said their organization decided not to buy at least one software product or feature. They could build it in house with agentic coding tools.
Developers are trying them, and they are worried
The OECD paper reads the 2025 Stack Overflow Developer Survey. That survey received more than 49,000 responses from 177 countries and covered 62 questions. Stack Overflow, as quoted by the OECD, defines this AI as autonomous software that can operate with minimal or no direct human intervention, using AI techniques.
About half of the people who answered that question are already using this kind of AI, or plan to. 38% have no plans to adopt them. The OECD calculated those shares from people who answered the question. That group was 31,890 people.
The OECD says the vast majority of developers in that survey highlight a need to strengthen security, privacy, and accuracy.
Long tasks were still hard in a 2025 study
On March 19, 2025, the research group METR published a way to measure this AI by task length. Length means how long a skilled person takes, not how long the model runs.
In that post, METR said models of that moment had almost 100% success rate on tasks that take a person less than 4 minutes. They succeeded less than 10% of the time on tasks that take more than about 4 hours.
METR also said the best of this AI then could not carry out a substantive project by themselves. They could not reliably handle even relatively low skill computer work, such as remote executive assistance.
METR warns, on that same page, that some text and figures in the post are out of date. Those claims reflect the data at publication. METR says it is no longer updating that chart with new models. Treat this section as a March 2025 snapshot, not a live score.
Sources
Opened and checked on November 10, 2026.
- OECD, Luis Aranda and Kasumi Sugimoto, The agentic AI landscape and its conceptual foundations, OECD Artificial Intelligence Papers, No. 56, February 2026. DOI: 10.1787/396cf758-en.
- Shana Lynch, Stanford Institute for Human-Centered AI, Inside the AI Index: 12 Takeaways from the 2026 Report, April 13, 2026.
- McKinsey, The state of AI in 2026: On the road to ROI, Global Survey, August 25, 2026.
- Thomas Kwa, Ben West, Joel Becker, and others, METR, Measuring AI Ability to Complete Long Software Tasks, March 19, 2025.