From Chatbot to Coworker: Gemini 3 Ushers in the Agent Era

Mollick showcases how Gemini 3 moves beyond chat to agentic work: planning, coding, browsing, and building deployable products with human oversight. In research tasks, it cleaned data, ran analyses, and drafted a paper with an original NLP metric, showing strengths akin to a solid PhD student alongside fixable judgment issues. He concludes the era is shifting from chatbots to digital coworkers, with humans acting as managers and stewards of quality and safety.

Key Points

• AI has evolved from text chat to agentic systems that plan, code, and autonomously operate computers, making them general-purpose digital workers.
• Google’s Gemini 3 with Antigravity can read local files, browse, build and test websites, and handle deployment, while checking in for permissions and revisions.
• In a research test, Gemini 3 cleaned messy legacy data, devised hypotheses, ran analyses, and wrote a 14-page paper, including an original NLP-based novelty metric.
• The model’s errors resembled human judgment issues (method choices, theory stretch) rather than classic hallucinations, and improved with managerial guidance.
• The human role is shifting from policing outputs to managing and directing capable AI coworkers; however, giving agents system access carries security and data risks.

Sentiment

The overall sentiment is notably skeptical, though not dismissive. A substantial faction pushes back against the article's optimistic framing, arguing that AI demos consistently overstate capabilities, that outputs are unreliable in production, and that the agent era narrative obscures fundamental architectural limitations. However, a meaningful minority of commenters share concrete, positive experiences with AI-assisted coding and research. The discussion reflects a community deeply divided between cautious pragmatists who see real but limited utility in AI tools, and critics who view the hype as dangerously misleading. The energy leans more toward 'this is oversold and we need to be honest about limitations' than toward celebration.

In Agreement

• AI coding agents have crossed a threshold where shipping entire projects without manually writing code is now feasible, as demonstrated by concrete examples of shipped products.
• Gemini 3 specifically represents a qualitative leap, introducing concepts into conversations unprompted and serving as a genuine intellectual collaborator rather than just a code generator.
• Agentic AI's ability to plan, experiment, evaluate results, and replan is genuinely impressive and more meaningful than simple knowledge retrieval.
• Individual developer productivity can multiply dramatically, allowing one person to accomplish in an afternoon what previously required a team over multiple sprints.
• AI tools are highly useful as collaborative sounding boards for writing, research, and problem-solving, even if their raw output is not directly copy-pasteable.
• By most measurable metrics, models have improved significantly over three years, and the right approach is to use AI as a force multiplier under human direction.
• Companies have years of backlog they lack capacity to address, so AI-augmented productivity opens new markets and possibilities rather than simply eliminating jobs.

Opposed

• The article follows a pattern where AI demos impress in areas the author cannot deeply evaluate, while admitted shortcomings in familiar areas suggest the impressive results are illusory (Gell-Mann amnesia effect).
• AI-generated code looks plausible at first glance but consistently contains bugs, security problems, and architectural issues that make it unshippable in complex production systems.
• Transformer-based LLMs are architecturally incapable of true novelty since they select the most probable next token, reproducing existing patterns rather than generating new ideas.
• The research paper example is disingenuous: a cherry-picked, non-peer-reviewed analysis that demonstrates cargo-cult science and contributes to an epidemic of fraudulent AI-generated academic papers.
• LLMs have fundamentally stalled since ChatGPT's release; improvements are in tooling and scaffolding, not in the models themselves.
• Giving AI agents system access creates serious security risks, and the industry has set security practices back significantly by normalizing the sending of sensitive data to third-party companies.
• In real enterprise environments, writing code represents only a small fraction of the work, so faster coding has minimal practical impact on overall productivity.
• The 'you're using it wrong' response to AI criticism is itself a red flag, suggesting the technology has fundamental usability and reliability problems that apologists deflect rather than address.