𝐖𝐞 𝐁𝐮𝐢𝐥𝐭 𝐏𝐨𝐞𝐭𝐬 𝐚𝐧𝐝 𝐀𝐬𝐤𝐞𝐝 𝐓𝐡𝐞𝐦 𝐭𝐨 𝐃𝐨 𝐎𝐮𝐫 𝐓𝐚𝐱𝐞𝐬 Sam Altman promised 2025 would be the year AI agents “join the workforce.” It wasn’t. The New Yorker just autopsied why. Two insights worth your time: First, the architecture mismatch. LLMs are pattern-matching engines. We trained them to predict the next word, not to model cause and effect. Then we asked them to book hotels, navigate websites, and complete multi-step tasks requiring actual reasoning. One demo tried planning a road trip to all 30 MLB stadiums. It included a stop in the middle of the Gulf of Mexico. That’s what happens when you force probabilistic text generators to do deterministic work. Second, the feedback loop problem. Coding agents actually work. GitHub Copilot, Cursor, Replit; they ship real value. Why? Because code has binary feedback. It compiles or it doesn’t. The agent gets immediate, clear signals about success or failure. But “book me a good hotel” has no such loop. Good for who? Measured how? The agent generates a plan with 18 sub-steps, each requiring judgment calls on undefined weights and preferences. One wrong move at step 4 and you’re sleeping in a hostel. No feedback mechanism can save you when the task itself is squishy. The tasks AI handles well aren’t the ones we thought. It’s not about complexity. It’s about whether the environment gives clear, fast feedback. Coding: binary signals, immediate results. Hotel booking: subjective goals, delayed feedback, no clear win condition. We optimized for the appearance of intelligence over the substance of capability. Built systems that write fluently about any topic but can’t reliably accomplish a single real-world task. The article quotes Andrej Karpathy, OpenAI co-founder: agents are “cognitively lacking” and “it’s just not working.” Even Altman quietly backed off in an internal memo. OpenAI is deemphasizing agents to focus on its core chatbot. Tasks with tight feedback loops get automated. Tasks requiring judgment in ambiguous contexts stay human. That’s not a technology limitation for 2025. That’s an architecture reality. Source: Cal Newport, The New Yorker, “Why A.I. Didn’t Transform Our Lives in 2025” https://lnkd.in/eZHuR7e3