When Did Practical AI Agents Become Possible?

Ethan Mollick says practical agents were not possible until models reached Opus 4.6 level or 4.5 level around November and December of last year. Agents became practical at that point. Co-Intelligence, his last book, was about the chatbot era because that was as good as models were. The newer book is about the agentic era because that is where models are today.

Product formats are shifting for three reasons at once. The study of innovation calls the first an era of ferment: nobody knows the right form, so ideas churn and get copied. OpenClaw came out in January as personal agents, and the form took root. Muse and Dot are claw-likes. Second, vibe coding and agents sit in different categories. Vibe coding is a human activity. Agents are a capability of AI systems, so they are not competing product lines. In companies he has talked to, vibe coding was never the planned way to use AI systems, yet vibe coding has taken off. He vibe codes work that would have required a coder before. People build games and very serious projects, so vibe coding has not gone away.

The most important cause is the capability curve. Moving up an exponential ability curve changes which products are possible. Competing groups are all trying to claim the idea of AI products while real capability gain shows up at the same time.