From ChatGPT to Agent-Native Everything: Where AI Is Headed Next
From Prompting to Orchestrating: The Future of Agent-Native Everything
From the humble Transformer to the modern agent, AI has evolved rapidly. Looking closely at things like the OpenHermes agent, OpenClaw, and the rise of coding agents in general, you can notice small shifts piling up to create something massive. To understand where AI is going, we have to look back at the history from our own perspective: the launch of ChatGPT, o1 Pro, Gemini 2.0 Flash, DeepSeek R1, Gemini 2.5 Pro, Cursor, Windsurf, Trae, OpenCode, GPT-5.5, and Pi. AI has progressed from chatting, to thinking, and now to orchestrating. The industry is moving forward into the agent side faster than ever, and people are noticing. Moving from simple prompting to complex loops, people are realizing that with enough time and money, you can create anything. That app that has been stuck in your head for five years? Or a brand-new idea to challenge existing systems? With enough tokens, the sky is the floor, not the limit. So, where exactly is AI going? Let's break down the shifts.
1. Computer Use: The End of Tech Support
Recently, OpenAI has been showing us a future of computer use where models can interact with your laptop even when it's locked. That is cool on its own, but why does it matter if computer use already technically exists? Because of what it unlocks for everyone else. Imagine helping your grandmother or a less tech-savvy friend. They might not know how to fix a system issue, but an model like GPT-5.5 will know exactly how to do it. Computer use for everyday consumers will mean the end of basic tech support. We will still need people to manage large fleets of computers, but fixing a driver issue? That’s over. On mobile, computer use will completely reinvent the smartphone. Imagine you build an app; the model installs it on your phone, tests it, makes UI changes, and deploys bug fixes all while you are asleep. The model is just using your computer and phone as a tool. Smartphone application building will become incredibly fluid, and customization will be seamless (that is, if we even still build traditional apps at all).
2. Skills over Traditional Apps
The new question every investor is asking themselves before backing a company comes down to two things: What is your edge factor, and can your startup be replaced by a simple markdown file? What are "skills"? Skills are essentially markdown files that give an AI agent context and instructions. With tools like GStack and Grill Me, people see that they can go much further with this concept. Instead of building a massive app, you can give an agent a set of instructions. If an app's entire job is just taking data from an API and transforming it into another form, it is cooked. But why fight it?
3. Agent-Native Everything
The future of software involves skills and MCP (Model Context Protocol) servers for your everyday tools. Instead of building applications exclusively for human eyes, apps won't just have an AI chatbot slapped on top. Instead, MCP or Skills + CLI will become the dominant way to build. This shifts the economic model. Instead of paying for a twenty-dollar API token subscription to twenty different services, users will rely on their core orchestration subscriptions (like ChatGPT or Claude Code) and let their personal agents pull data from their other apps. If a user pays for a high-tier subscription, they get massive token limits. They will just let their personal agents handle the heavy lifting across applications instead of paying every separate software company an AI premium. Furthermore, Agent SEO will become more important than regular SEO. If someone asks their agent to find them a restaurant, a park, or a backend platform, how likely is the agent to pick your product over a competitor? Optimizing for agent discovery will become critical in the near future.
4. Small is Better (The Agent Factory)
We are going to see a new wave of companies realize that a mixture of smaller models is where the real value of AI hides. You can often get 100x the value out of a small model compared to a massive one for specific tasks. Smaller models will be used for exploring, executing, and handling natural, isolated tasks. Meanwhile, the heavyweight models like Claude 4.8 Opus, GPT-5.5, and Fable 5 will act as the orchestrators. The software factory of the future will function like a traditional company:
- Tasks are broken down.
- Smaller models are assigned specific tasks and permissions.
- The larger model manages the permissions, reviews the work, and steps in for heavy token-generation and deep reasoning. We will also see a massive push for efficiency. If you have a small model that uses 3x more tokens than necessary, and a big model that costs 3x more per token but is highly efficient, the net cost of using both ends up being the same but the bigger model yields a much better result. The market will naturally gravitate toward the most mathematically efficient models.
5. Frontier Model Predictions
- Google Gemini 3 & 3.5: This series will be their defining moment. Either their tool-calling capabilities get permanently fixed, or they take the Mistral route, leaning further into open weights and specialized models. I don’t hate Google; I like Google and want them to win. To any Googler reading this: do whatever it takes to make Gemini excel, because the future of the company depends on it.
- GPT-5.6 & Fable: GPT-5.6 will be a good model, but it won't be quite at the "Fable" level. It will likely beat Opus 4.8 and maybe Gemini 3.5 Pro, but Fable is a different tier of intelligence entirely.
- xAI (Grok): Grok can absolutely catch up. Right now, some view it as mostly good for auto-tagging on X, but they are positioned to make a massive leap. Look at their compute capital, their data pipeline gotten form there acquiration of cursor, and what they can do with their infrastructure. They are in a prime spot to disrupt the leaderboard. for an example just look what cursor did with composer it 4 tries to catch up with frontier intelligence and know that same has the compute and captial to do something really cool.
- OpenAI Custom Silicon: OpenAI’s TPU initiatives will rival the likes of Google and Cerebras. Their partnership with Broadcom to develop custom ASICs is a clear sign. Whether it's a traditional TPU or specialized silicon, it will eventually power their upcoming hardware products and we know Jony Ive is cooking something interesting on the hardware front.
- On-Device Intelligence: This might be my boldest prediction yet, but I expect GPT-5 level intelligence running locally on consumer hardware by the end of this year or early next year. It will hit laptops first, and on phones, it will likely debut on Samsung, Apple, or Chinese flagship devices. Generally speaking, local models are getting cheaper and better so fast that prosumer hardware like a maxed-out MacBook Pro or a Framework desktop will easily run massive intelligence locally.
6. The Swing to Privacy
With the rapid acceleration of AI capabilities, I expect a massive swing toward absolute privacy. People will want private servers and self-hosted services for everything. However, because cyber capabilities and model response times are becoming so incredibly short, security demands will pull the market in both directions. Despite that friction, a deeply private, self-hosted local stack will be the ultimate goal for power users.
What's your take?
Let’s talk about it. Agree or disagree on the timeline for on-device frontier models? Drop a line on X at @JemoLife0213 and let me know your take.