Last Fall I traveled to MIT to watch my daughter play in the NCAA volleyball tournament. On the way, we passed signs for Lexington and Concord. AI agents were on my mind. There was a sudden buzz about AI agents and how they’re coming for our jobs. The image of Paul Revere came to my mind.
Instead of warning about the Redcoats stealing munition at Concord, Revere’s of today warn of AI agents stealing our jobs. Then new AI reasoning models released causing another rise in discussion. Like Lexington Green have the first shots been fired on our jobs with reasoning AI agents?
What is an AI agent?
Search interest in AI agents spiked in January. If you search AI agents Google returns 216 results. Reading through many of them there are probably half as many definitions. For simplicity, I will begin by quoting AI Marketing Institute’s Paul Roetzer, “An AI agent takes action to achieve goals.”
That doesn’t sound scary. What’s driving interest and fear is adding the word autonomous. Roetzer and co-founder Mike Kaput have created a helpful Human-to-Machine Scale that depicts 4 levels of AI autonomous action.
Marketing AI Institute’s Human-to-Machine Scale:
- Level 0 is all human.
- Level 1 is mostly human.
- Level 2 is half and half.
- Level 3 is mostly machine.
- Level 4 is all machine or full autonomy.
Full autonomy over complete jobs is certainly fear inducing! Large language model companies like OpenAI, Google, and SAAS companies integrating AI are promising increased autonomous action. Salesforce has even named their AI products Agentforce, which literally sounds like an army coming to take over our jobs! Put some red coats on them and my Paul Revere analogy really comes to life.
Every player in AI is going deep.
In September Google released a white paper “Agents” with little attention. Now, after the release of reasoning models, everyone including Venture Beat is analyzing it. In the paper, Google predicts AI agents will reason, plan, and take action. This includes interacting with external systems, making decisions, and completing tasks – AI agents acting on their own with deeper understanding.
OpenAI claims its new tool Deep Research can complete a detailed research report with references in “tens of minutes.” Something that may take a human many hours. Google’s DeepMind also has Deep Research, Perplexity has launched Deep Research, CoPilot now has Think Deeper, Grok3 has a Deep Search tool, and there’s the new Chinese company DeepSeek. Anthropic now has released what it is calling the first hybrid reasoning model. Claude 3.7 Sonnet can produce near-instant responses or extended step-by-step thinking that is made visible. The Redcoats are coming and they’re all in on deep thinking.
What is a reasoning model?
Google explains Gemini 2.0 Flash Thinking is “our enhanced reasoning model, capable of showing its thoughts to improve performance and explainability.” A definition for reasoning models may be even more difficult and contested than AI agents. This term returns 163 results in a Google search and perhaps just as many definitions.
For my definition of a reasoning model, I turn to Christopher Penn. In his “Introduction to Reasoning AI Models,” Penn explains, “AI – language models in particular – perform better the more they talk … The statistical nature of a language model is that the more talking there is, the more relevant words there are to correctly guess the next word.” Reasoning models slow down LLMs to consider more words through a process.
LLMs and reasoning models are not magic.
Penn further explains that good prompt engineering includes a chain of thought, reflection, and reward functions. Yet most people don’t use them, so reasoning models make the LLM do it automatically. I went back to MIT, not for volleyball, but for further help on this definition. The MIT Technology Review explains that these new models use using chain of thought and reinforcement learning through multiple steps.
An AI prompt framework, such as the one I created, will improve your results without reasoning. You also may not need a reasoning model for many tasks. Reasoning models cost more and use more energy. Experts like Trust Insights recommend slightly different prompting for reason models such as Problem, Relevant information, and Success Measures. Brooke Sellas of B Squared Media shared President of OpenAI Greg Brockman’s reasoning prompt of Goal, Return Format, Warnings, and Context Dump.
Many want a magical AI tool that does everything. In reality, different AI is better for different things. Penn explains generative AI is good with language, but for other tasks, traditional forms of AI like regression, classification, or even non-AI statistical models can be a better solution.
How we talk about AI matters.
Humans are attracted to the magic capabilities of AI. Folk tales like The Sorcerer’s Apprentice which you may know from Disney’s Fantasia, are about objects coming to life to do tasks for us. Reasoning models are said to have agentic behavior – the ability to make independent decisions in pursuit of a goal. Intentional or not, it sounds like angelic, bringing up mystical thoughts of angels and the supernatural.
Since the first post in my AI series, I’ve argued for maintaining human agency and keeping humans in the loop. Therefore, I want to be careful in how I talk about these new “reasoning” models that show us their “thinking.” I agree with Marc Watkin’s recent Substack “AI’s Illusion of Reason,” that the way we talk about these AI models matters.
An AI model that pauses before answering and shows the process it followed doesn’t mean it is thinking. It’s still a mathematical prediction machine. It doesn’t comprehend or understand what it is saying. Referring to ChatGPT or Gemini as it versus he or she (no matter the voice) matters.
What’s the difference between human and AI thinking?
I asked Google’s reasoning model Gemini 2.0 Flash the difference between human thinking and AI thinking. It said, “AI can perform tasks without truly understanding the underlying concepts or the implications of its actions. It operates based on learned patterns, not genuine comprehension.” Does this raise any concerns for you as we move toward fully autonomous AI agents?
Humans need to stay in the loop. Even then, you need a human who truly understands the subject, context, field, and/or discipline. AI presents its answers in a convincing, well-written manner – even when it’s wrong. Human expertise and discernment are needed. Power without understanding can lead to Sorcerer’s Apprentice syndrome. A small mistake with an unchecked autonomous agent could escalate quickly.
In a Guardian article, Andrew Rogoyski, a director at the Institute for People-Centred AI warns of people using responses by AI deep research verbatim without performing checks on what was produced. Rogoyski says, “There’s a fundamental problem with knowledge-intensive AIs and that is it’ll take a human many hours and a lot of work to check whether the machine’s analysis is good.”
Let’s make sure 2025 is not like 1984.
I recently got the 75th anniversary edition of George Orwell’s 1984. I hadn’t read it since high school. It was the inspiration behind Apple’s 1984 Super Bowl ad – an example of the right message at the right time. It may be a message we need again.
AI isn’t right all the time and right for everything. It’s confident and convincing even when it’s wrong. No matter how magical AI’s “thinking” seems, we must think on our own. As AI agents and reasoning models advance, discernment is needed, not unthinking acceptance.
The 250th anniversary of Paul Revere’s ride and the “Shot heard ‘round the world” is in April this year. Will AI agents and reasoning models be a revolution in jobs in 2025? In my next post, “How Will AI Agents Impact Marketing Communications Jobs & Education? See Google’s AI Reasoning Model’s “Thoughts” And My Own” I take a deep dive into how AI may impact marketing and communications jobs and education. What’s your excitement or fear about AI agents and reasoning models?
This Was Human Created Content!
