We went from basic LLMs to prompt engineering to context engineering, and now everyone is talking about agent engineering. Each step felt like a revolution at the time, but looking back, the progression was inevitable. The question is: what comes next, and have we finally hit the ceiling of what software alone can do?

The evolution of talking to AI

It started simply enough. Early large language models were impressive but clumsy. You typed something in, you got something back, and if it was wrong, you tried again with different words. That was the era of basic LLM usage, a glorified guessing game with a very smart autocomplete. Then came prompt engineering, the realization that how you asked mattered enormously. Researchers and practitioners discovered that carefully structured prompts, few-shot examples, and chain-of-thought reasoning could dramatically improve output quality. Suddenly there was a whole discipline around crafting the right input. Multi-shot prompting gave models examples to learn from in-context. Chain-of-thought prompting encouraged models to reason step by step rather than jump straight to an answer. But prompt engineering had limits. You were still working within a single interaction, manually crafting every instruction. The context window was a fixed resource, and you had to be clever about what went into it.

Enter context engineering

By mid-2025, a new term gained traction: context engineering. As Anthropic described it, building with language models became "less about finding the right words and phrases for your prompts, and more about answering the broader question of what configuration of context is most likely to generate the model's desired behavior." Context engineering moved beyond crafting individual prompts. It focused on the entire information architecture surrounding a model, including retrieval-augmented generation (RAG), memory systems, structured templates, tool-aware prompts, and scratchpads for intermediate reasoning. Instead of prompt engineers, the industry started talking about context architects who design data flow, retrieval logic, and prompt orchestration for complex systems. The insight was simple but powerful: models may be getting stronger, faster, and cheaper, but no amount of raw capability replaces the need for well-structured memory, environment, and feedback. As the team at Manus put it, "how you shape the context ultimately defines how your agent behaves."

So what is agent engineering?

Agent engineering is the next step. LangChain defines it as "the iterative process of harnessing LLMs into reliable systems." But that undersells what is actually happening. Agent engineering is a full discipline that combines three pillars: Product thinking defines the scope and shapes agent behavior. This means writing detailed prompts (often hundreds or thousands of lines), deeply understanding the job the agent needs to do, and defining evaluations that test whether the agent performs as intended. Engineering builds the infrastructure for production readiness. This includes writing tools for agents to use, developing UI/UX for agent interactions with streaming and interrupt handling, and creating robust runtimes that handle durable execution, human-in-the-loop pauses, and memory management. Data science measures and improves agent performance over time through evaluations, A/B testing, monitoring, and reliability metrics. The key difference from everything that came before is that agent engineering treats AI systems as autonomous actors rather than sophisticated text completers. Agents don't just respond to prompts. They plan, use tools, maintain state across interactions, and take actions in the real world.

The rise of agents

2025 was widely dubbed "the year of the agent." IBM and Morning Consult surveyed 1,000 developers building AI applications for enterprise, and 99% said they were exploring or developing AI agents. The hype was real, but so was the reality check. As Yi Zhou wrote on Medium, "the mistake was never believing agents would matter. The mistake was believing they would arrive fully formed." What many labeled the year of the agent was always going to be something more like the decade of the agent. Building real engineering disciplines takes time. Civil engineering did not emerge from stronger materials alone. Software engineering did not emerge from faster hardware. Autonomous systems will not emerge from better language models by themselves. Each leap requires a shift from capability to discipline, from "we can build this" to "we know how to build this safely, repeatedly, and at scale." That shift is still underway, but the trajectory is clear. Agentic AI is projected to power everything from autonomous research assistants to advanced customer service solutions, self-healing data pipelines, and multi-agent ecosystems that coordinate like teams of specialists.

Agents might be the final form of AI software

Here is a provocative thought: agents may represent the final major paradigm shift in AI software. We have gone from raw models to engineered prompts to engineered context to engineered agents. Each step added a layer of sophistication around the same core technology, large language models. Agents feel like the natural endpoint of that progression because they encompass everything that came before. An agent still needs good prompts, good context, and good models. It just wraps them in autonomy, tool use, and persistent state. If that is true, then the next frontier is not another software abstraction. It is hardware.

The future is physical

The signs are everywhere. We are witnessing the largest capital investment in the history of humanity, and it is flowing into physical infrastructure. In 2025 alone, hyperscalers invested $325 billion in AI data centers and infrastructure. By 2026, Alphabet, Amazon, Meta, and Microsoft collectively plan to spend over $600 billion in capital expenditure, with more than 75% directed toward AI infrastructure. McKinsey estimates that a total of $6.7 trillion will be spent globally on AI infrastructure by 2030. To put that in perspective, the entire Apollo Program cost $300 billion in inflation-adjusted dollars. The Manhattan Project cost $30 billion. No previous human undertaking matches the scale of what is being built right now to power AI. And the hardware story extends far beyond data centers. Humanoid robots are having their moment.

The robot race

Tesla Optimus is perhaps the most high-profile humanoid robot project. Tesla is set to unveil Optimus Gen 3, with initial production expected later in 2026 and plans for public sales by the second half of 2027. Wall Street analyst Dan Ives projects Tesla could reach a $2 trillion market cap by end of 2026 and $3 trillion by end of 2027, driven largely by full self-driving and robotics growth. But Tesla is far from alone. Figure AI raised over $1 billion at a staggering $39 billion valuation. Physical Intelligence secured $400 million from backers including Jeff Bezos. SoftBank agreed to acquire ABB's robotics division for $5.375 billion. Chinese companies like XPeng, UBTech Robotics (with their Walker S2), and Estun Automation (with the Codroid 02) are pushing hard into the space. South Korea's Rainbow Robotics developed the RB-Y1 semi-humanoid. The global robotics funding surpassed $10.3 billion in 2025, the highest since 2021. The broader global robotics market is expected to grow from $76 billion in 2023 to $218 billion by 2030, according to GlobalData. The surgical robotics market alone is projected to reach $14.45 billion in 2026.

Hardware always wins

There is a pattern worth noticing. Every major computing revolution eventually becomes a hardware story. The internet era needed fiber optic cables, routers, and server farms. Mobile needed chips small enough and efficient enough for pockets. Cloud computing needed massive data centers. And AI needs GPUs, custom accelerators, nuclear power plants, and eventually physical robots to operate in the real world. Nvidia has dominated the GPU market powering AI training and inference. But the landscape is shifting. Broadcom anticipates its AI revenue doubling, with custom AI processor shipments projected to increase 44% in 2026 compared to 16% growth in GPU shipments. Companies are increasingly investing in custom silicon, and data center operators are upgrading power distribution from 208V to 415V to handle the enormous energy demands. The energy requirements alone are staggering. The AI industry is driving a resurgence in nuclear energy interest, with 71 reactors under construction globally and Bloomberg Intelligence predicting U.S. nuclear capacity may grow by 63% by 2050. Global data center infrastructure spending is on course to surpass $1 trillion annually by 2030.

What this means

The AI story has two chapters. The first chapter, the one we have been living through, is about software: better models, better prompts, better context, better agents. That chapter is maturing rapidly. The second chapter is about hardware: the physical infrastructure to run these agents at scale, and the physical robots that let agents interact with the real world. That chapter is just beginning, and the investment numbers suggest it will be far larger than anything we have seen before. Agent engineering is not just a buzzword. It is a real discipline that is still being defined, one that combines product thinking, engineering, and data science into something genuinely new. But agents running on screens are only half the story. The other half is agents running on legs, arms, and wheels, powered by trillion-dollar infrastructure that is being built right now. The software got smart. Now the hardware needs to catch up.

References