AI Agents Execute Economically Meaningful Work Through Reasoning

Key Quotes

"The gap between how most people use ai today and what it's actually capable of is mountainous that's because many business leaders are driving by only looking in the rear view mirror many enterprises are primarily still using ai to polish their blog posts or as a google search replacement and that's one of the main reasons i think 2025 wasn't quite the breakout year for ai that many thought it might be but that's also because the technology in itself in the first half of the year needed a lot of elbow grease and duct tape it's not the case in 2026 today's models are so good you can sneeze and accidentally create a million dollar app i mean with little experience and a few clicks you can send a swarm of capable agents to go accomplish real work for you"

The speaker, Jordan Wolsen, argues that the widespread underutilization of AI stems from a limited perspective among business leaders, who are still treating AI as a tool for basic tasks like content polishing or search replacement. He contrasts this with the advanced capabilities available in 2026, where AI can generate complex applications and deploy agents for significant work with minimal user input. This highlights a disconnect between AI's potential and its current adoption.

"Number one is reasoning by default so large language models are now a natural thought partner and they were just kind of a fun tool maybe last year for a lot of people Number two is agentic scaffolding so we've gone from passive knowledge to active execution with tools Number three is frictionless data integration so rag pipelines if i'm being honest i'm bearish right i'm very bullish on what most front end large language models offer now is one click versions to bring your company's dynamic context to the front Number four is exponential task endurance so going from what large language models were capable of maybe a year a year and a half ago short sprints to now long term projects and then Number five and this is last but definitely not least is economically meaningful work yeah that large language models are no longer about writing better blog posts or replacing google they are about doing the actual work that you're doing now"

Jordan Wolsen outlines five key pillars driving AI acceleration in 2026: reasoning capabilities, agentic scaffolding for active execution, simplified data integration, extended task endurance, and the ability to perform economically meaningful work. He posits that these advancements shift AI from a simple tool to a capable partner that can handle complex, real-world tasks, moving beyond basic content generation or search functions. This signifies a fundamental change in how AI will be integrated into business operations.

"The gap between how most people use ai today and what it's actually capable of is mountainous that's because many business leaders are driving by only looking in the rear view mirror many enterprises are primarily still using ai to polish their blog posts or as a google search replacement and that's one of the main reasons i think 2025 wasn't quite the breakout year for ai that many thought it might be but that's also because the technology in itself in the first half of the year needed a lot of elbow grease and duct tape it's not the case in 2026 today's models are so good you can sneeze and accidentally create a million dollar app i mean with little experience and a few clicks you can send a swarm of capable agents to go accomplish real work for you"

Jordan Wolsen explains that the significant difference between AI's current usage and its true potential is due to business leaders' outdated perspectives, viewing AI as merely a tool for tasks like blog writing or search. He attributes the slower adoption in 2025 partly to the technology's immaturity, requiring extensive effort to implement. Wolsen emphasizes that in 2026, AI models are so advanced that they can facilitate the creation of valuable applications and deploy agents for substantial work with minimal user effort.

"What's interesting to note is a stat that's not talked about a lot last year sam altman said that only 7 of queries involved reasoning models and that to me is absolutely nutty all right so if you are uh kind of newish or not super technical let me explain very briefly kind of this big shift that happened in 2025 and it was a slow shift right because maybe a lot of the a lot of how people view large language models is the original chatgpt right a friendly chatbot that's you know really fun can spit out you know large blocks of text and sometimes hallucinates and uh it's not really what large language models are today i'd like to say that there's almost a line in the sand and to me that line it's not between uh you know normal ai and and agents it's actually between you know kind of quote unquote old school transformer models versus new reasoning models and i think that this has turned large language models from something that humans have to put a lot of work into into being a true agentic partner because the ability to reason is an absolute game changer"

Jordan Wolsen highlights the low adoption rate of reasoning models in AI queries, citing Sam Altman's statistic of only 7%. He explains that this is a critical shift from earlier AI models, which were perceived as simple chatbots capable of generating text but prone to errors. Wolsen argues that the introduction of reasoning capabilities has transformed Large Language Models (LLMs) from tools requiring significant human effort into true "agentic partners," fundamentally changing their utility and potential.

"The benchmark consists of more than 1300 specialized tasks across 44 distinct occupations right the in the nine in the nine sectors that contribute most to the us's gdp and then there's blind expert grading so outputs outputs are graded via blind pair pairwise comparison by human industry professionals who have an average of 14 years of experience and then they judge whether the ai's work is superior equal or inferior to a human expert's attempt so it's kind of just like a blind you know a blind taste test on real work that all of us do right literally all of us and well here's what they found they found that the model gpt52 and i believe they used the thinking version achieved a 74 win rate right whereas the model from just a couple months prior gpt5 that was kind of not well received right it got a 38 win rate so it almost doubled its win rate of creating economically valuable work"

Jordan Wolsen describes the GDP-Val benchmark, which assesses AI models on over 1300 real-world tasks across various occupations and economic sectors. He explains that the evaluation involves blind comparisons by human experts with an average of 14 years of experience, judging AI output against human performance. Wolsen points out that GPT-5.2, using its thinking capabilities, achieved a 74% win rate, a significant improvement over the previous GPT-5 model's 38% win rate in generating economically valuable work. This demonstrates a substantial leap in AI's ability to perform complex, real-world tasks.

"No large language models if you know what you're doing they instantly connect to your business