Gemini 3: Please welcome your new co-worker

I have a confession: I was getting bored.

For the last eighteen months, the AI updates have felt relentless yet weirdly stagnant. Slightly faster tokens, slightly larger context windows, slightly better benchmarks that only researchers care about. We were stuck in the “Chatbot Era”—a phase where we marvelled at a machine that could talk, even if it frequently lied or forgot what we were talking about three messages prior.

Thanks for reading Cowritten.ai! Subscribe for free to receive new posts and support my work.

That era ended this morning.

With the release of Google’s Gemini 3 and its new agentic workspace, Antigravity, the paradigm has shifted. We are no longer typing into a void and hoping for a clever response. We are now logging in to manage a junior employee.

If you work in strategy, content, or creative direction, you might be tempted to glaze over the technical details of “agents,” “Generative UI,” and “coding environments.” Don’t. This is not a developer update. This is a rewriting of your job description.

Here is why Gemini 3 matters, and why you need to stop treating AI like a search engine and start treating it like a co-worker.

Beyond the Chatbox: The Rise of Agents

The fundamental frustration with previous models — even the good ones like GPT-4o (or 5 or even 5.1) or Claude 4.5 Sonnet — was their isolation. You asked for a strategy; they gave you text. You asked for a website; they gave you code snippets you had to assemble yourself. They were brains in a jar.

Gemini 3 has been given hands.

The leap here isn’t just raw intelligence (though the benchmarks are impressive); it is agency. Agentic AI doesn’t just describe a task; it executes it. It plans, it browses the web to verify facts, it writes code, it runs that code to test if it works, and then it fixes its own errors.

In my testing over the last 24 hours, the difference in “feel” is stark. When I asked previous models to “analyse the competitive landscape for a fintech brand,” I got a bulleted list of generic observations.

When I asked Gemini 3, it didn’t just answer. It went away, browsed live sites, pulled recent pricing changes, structured the data, and flagged a contradiction in a competitor’s press release. It feels less like a magic 8-ball and more like sending a brief to a smart, eager analyst who sits three desks away.

Generative UI: The Death of “Text-Only”

Perhaps the most immediate change you will notice isn’t just what the AI says, but how it displays it. We are seeing the debut of Generative UI.

Until now, our interaction with AI has been brutally Skeuomorphic: we type in a chat box, and it replies in a chat bubble. It’s text-in, text-out. But text isn’t always the best way to digest information.

With Gemini 3’s Generative UI capabilities, the model doesn’t just retrieve content; it designs a custom interface to present it.

If you ask for a comparison of electric cars, it won’t just write a paragraph. It might spin up an interactive comparison table with sortable columns. If you are brainstorming brand colours, it won’t list hex codes; it will generate a colour palette picker you can adjust in real-time.

For creative teams, this is a massive unlock. We are moving from “reading answers” to “using tools” that didn’t exist five seconds ago. The AI is effectively coding a micro-app on the fly, specifically tailored to answer your specific question. It is the ultimate responsive design — responsive not just to screen size, but to intent.

Antigravity and “Vibe Coding”

This ability to generate interfaces leads us to Google’s new standalone tool: Antigravity. On paper, it’s a code editor. If you are a writer or strategist, your instinct is to ignore it because you “don’t code.”

This is a mistake.

Antigravity is the first mainstream tool from Google designed for “vibe coding”. This means building software not by writing syntax, but by describing the outcome you want.

I tested this by attempting to build a simple but useful tool: a dedicated ContentOps dashboard that automates the tedious parts of SEO research.

I didn’t write a single line of Python or React. Instead, I gave it a paragraph of natural language — essentially a product brief:

“I want to prototype a ContentOps tool that focuses only on one simple content marketing use case: user inputs a keyword or long-tail query, the tool goes off to run an automated workflow that searches Google SERP, analyses top ranking pages, scrapes them, returns analysis of search intent and content gaps, then uses an LLM to create a brief, an outline and a draft based on the findings. Consider this an MVP so don’t over-engineer it, but make sure the UI is slick, modern and functional.”

In minutes, Antigravity built a functioning application (with mock APIs for now, so it doesn’t actually work yet, but that’s the easy part).

It wired up the search logic, created distinct tabs for ‘Analysis’, ‘Brief’, and ‘Outline’, and styled it in a dark-mode interface that looks like a SaaS product I’d pay $29/month for. It understood that “Content Gaps” needed to be a bulleted list for readability. It understood that “Search Intent” required a summary card.

But isn’t this just another coding assistant?

You might rightly ask how this differs from existing tools like Lovable, Claude Code, or OpenAI Codex.

The distinction is in the autonomy. Tools like Lovable are brilliant at generating beautiful UIs, and tools like Claude Code are exceptional at logic, but they often require you to act as the “integrator” — pasting errors back and forth or setting up the environment. They make coders faster.

Antigravity feels different because it manages the entire stack. It handles the “glue” — the messy API connections, the error handling, the deployment environment — that usually stops non-technical creatives dead in their tracks. It generates the code, yes, but it also runs it, critiques it, tests it (in its own Chrome browser), and fixes it.

I described the function and the vibe; the machine handled the execution.

For creative directors and content leads, this destroys the barrier between “idea” and “prototype.” You no longer need to wait for an engineering sprint to test a concept. You can vibe-code a working tool in a morning, use it to validate your workflow, and then decide if it’s worth building for real.

Google Search is now a Research Department

While Antigravity handles creation, the new Gemini 3 Search AI Mode transforms how we gather information.

We have all spent years “Googling” — a manual process of hunting, pecking, clicking blue links, and synthesising information in our heads. Search AI Mode attempts to automate the synthesis.

Rather than just finding pages, it reasons across them. If you ask, “Plan a 3-day itinerary for Tokyo focusing on vintage denim shops, checking their opening hours against my arrival time,” it acts as a travel agent. It verifies the data, cross-references logical constraints, and presents a plan.

For strategists, this shifts “Googling” from a task of retrieval to a task of review. The AI does the heavy lifting of the “Deep Research” — reading the papers, checking the forums, aggregating the reviews — and presents a briefing document.

However, a warning: Trust, but verify. While Gemini 3 is better at citing sources, the risk of “plausible drift” remains. It might not hallucinate a fact, but it might misinterpret a nuance. Your role as the human expert is to spot check the logic, not just accept the summary.

The “Co-worker” Mindset

Ethan Mollick, a professor at Wharton who studies AI, put it perfectly in his recent analysis: we are moving from a world where humans fix AI mistakes to a world where humans direct AI work.

This requires a profound shift in how we show up.

In the Chatbot Era, our job was essentially editorial. We prompt, we read, we correct the hallucinations, we rewrite the robotic prose. It was low-leverage work.

In the Agentic Era, our job is managerial.

When you work with Gemini 3, you have to treat it like a talented junior staff member. You don’t just say “write a blog post.” You provide the context, the constraints, the resources, and the definition of success. And then — crucially — you let it go and do the work, checking in only to course-correct.

The “Manager’s Dilemma”

This brings us to a new bottleneck. If the AI can do the execution (the coding, the researching, the drafting), the value of a human shifts entirely to Intent and Taste.

Intent: Can you articulate exactly why we are doing this project and what “good” looks like? (See my previous notes on the specification bottleneck).

Taste: Can you look at the output — whether it’s a Generative UI graph or a drafted strategy — and instantly recognise if it meets the standard?

If you cannot define the destination, a fast car (or a fast AI) just gets you lost quicker.

Practical Steps for the Non-Technical

If you are reading this and feeling overwhelmed, don’t be. You don’t need to learn Python today. But you do need to change your habits.

Here is what I suggest:

Stop “Chatting”, Start “Assigning”: When you open Gemini 3, don’t ask a question. Assign a task. Give it a role. “You are a research assistant. Your goal is to find X. Here are the constraints. Create a plan first, then execute.”

Ask for the Interface: Don’t settle for text. If you are analysing data, ask Gemini: “Can you build me a calculator for this?” or “Make this into an interactive chart.” Push the Generative UI to see what tools it can build for you.

Audit, Don’t Edit: When the AI returns work, stop fixing commas immediately. Look at the logic. Did it understand the assignment? If not, don’t fix the output — fix the prompt. Train the employee, don’t just fix the work.

A Final Reflection

There is a seductive danger here. As these agents become competent co-workers, it will be easy to abdicate responsibility. To let them run the research, build the app, and draft the copy, while we nod along.

But remember Cowritten.ai’s manifesto: Human expertise in the loop.

Machines can draft, code, and simulate, but only humans can decide what matters. Gemini 3 is a mirror. If your strategy is vague, it will build a high-fidelity, functioning monument to that vagueness.

The era of the chatbot is over. The era of the co-worker has begun.

Be a good manager.

Thanks for reading Cowritten.ai! Subscribe for free to receive new posts and support my work.