The next wave: AI agents, physical AI, and what comes after
This series started with frustration — the gap between what AI can supposedly do and what most people are actually getting from it. We've spent eleven posts closing that gap: understanding how AI thin...
The next wave: AI agents, physical AI, and what comes after
Part 12 of the "Build with AI" series — The Final Post
This series started with frustration — the gap between what AI can supposedly do and what most people are actually getting from it. We've spent eleven posts closing that gap: understanding how AI thinks, building with data and prompts, automating with agents, shipping to production, and building responsibly.
Now we look forward.
Because the gap isn't standing still. The tools are evolving faster than almost any technology in history. And the next wave isn't a refinement of what we've been discussing — it's a different category of capability entirely.
This post is about what's coming. Not to predict the future with false precision, but to give you a map of the territory so you can position yourself intelligently. The builders who understand what's happening now will have a real advantage as the next wave arrives.
Where we are right now
Before looking forward, it's worth anchoring on what's actually true today — not hype, not speculation.
Agentic AI is operational. The multi-step, tool-using, decision-making agents we discussed in Post 7 are not experimental. They're being deployed at scale in enterprise workflows, customer service, software development, sales, and research. Anthropic's Claude Code, OpenAI's Operator, and Google's Gemini are all running autonomous multi-step tasks routinely.
AI employees are real, in narrow domains. Companies like Artisan, 11x, and Relevance AI have deployed AI SDRs — sales development representatives — that prospect, research, write outreach, follow up, and book meetings autonomously. They're not perfect. They break on edge cases. But they're running in production for real companies generating real revenue.
The software development loop is closing. As of early 2026, Claude Code is widely considered the best AI coding assistant, handling full codebases, multi-file edits, and agentic coding tasks that would have required senior engineers a year ago. The gap between "describing software" and "having software" has collapsed to hours in many domains.
This is the baseline. Everything that follows is what's happening next.
Wave 1 already breaking: autonomous agent networks
The first shift already underway is the move from single agents to networks of agents — systems where multiple AI agents work together, each specialized, coordinating toward shared goals.
During the next decade, the intersection of agentic AI systems with physical AI robotic systems will result in robots whose "brains" are agentic AIs — able to adapt to new environments, plan multi-step tasks, recover from failure, and operate under uncertainty.
But in pure software, it's already here. The pattern: an orchestrator agent receives a goal, breaks it into subtasks, assigns each to a specialist agent, coordinates the results, and delivers the output. The specialist agents might handle research, writing, fact-checking, formatting, and quality review — each optimized for its task, running in parallel.
What this makes possible in practice:
Research at unprecedented speed. A swarm that simultaneously searches dozens of sources, extracts structured data from each, cross-references for consistency, and synthesizes into a report — completing in minutes what would take a human researcher days.
Software teams in miniature. Agent networks that handle the full software development loop: requirements analysis, architecture planning, implementation, testing, debugging, documentation, and deployment — with a human reviewing and approving at key checkpoints.
Business process automation at depth. Not just routing emails or filling forms, but handling the full lifecycle of a business process: intake, analysis, decision, action, follow-up, exception handling — with humans intervening only for genuinely novel situations.
If you're designing agent systems today, design them to be composable. Build agents that can call other agents. Build orchestration layers that can coordinate them. The systems that will scale are the ones built as networks, not monoliths.
Wave 2 building: physical AI
The most dramatic development of 2026 — and the one with the longest reach — is the convergence of AI intelligence with physical hardware. Physical AI.
At CES in Las Vegas, NVIDIA CEO Jensen Huang said the "ChatGPT moment for physical AI is here," marking an inflection point in the robotics space.
What does that actually mean?
For decades, robots have been scripted machines. They follow pre-programmed routines — precise, reliable within their defined parameters, useless outside them. A factory robot that welds the same joint 10,000 times a day is extraordinarily capable at that exact task and incapable of anything else.
The shift: AI-powered robots can now perceive their environment, understand natural language instructions, reason about what to do next, and adapt to situations they've never encountered before. They're not following a script — they're thinking.
The enabling technology is VLA models — Vision-Language-Action models that integrate visual perception (seeing the environment), language understanding (processing verbal commands), and action execution (moving in the physical world). VLA models enable robots to interpret visual inputs and language commands, transforming them into actionable tasks — replacing outdated scripted routines.
Who is building physical AI right now
NVIDIA's Isaac GR00T N1.6 is an open reasoning vision language action model, purpose-built for humanoid robots, that unlocks full body control and uses NVIDIA Cosmos Reason for better reasoning and contextual understanding. ABB Robotics, AGIBOT, Agility, Figure, KUKA, Universal Robots, and YASKAWA are among the leaders building on NVIDIA technology to deploy physical AI at scale.
Tesla's Optimus Gen 2 is designed to handle repetitive tasks, assist in manufacturing, and perform home automation functions — learning from real-world data. Boston Dynamics unveiled its all-new Electric Atlas at CES 2026, a high-performance, enterprise-grade humanoid designed for industrial tasks from material handling to order fulfillment. 1X has opened preorders for its NEO home humanoid, with first customer deliveries planned for 2026.
Chinese humanoid robot makers showed off their wares at Smart Factory & Automation World in Seoul in March 2026 — the entire COEX venue, 2,300 booths. AGIBOT has announced the rollout of its 10,000th humanoid robot, becoming one of the first companies in the industry to reach this milestone at scale.
Where physical AI is already working
The robotics field in 2026 has moved from experimental novelty to active pilot programs in warehouses and factories.
Amazon's warehouse fleet crossed 1 million robots, with its DeepFleet AI boosting travel efficiency by 10% across its network. Surgical robotics have reached 60% adoption in large hospitals, with robotic-assisted procedures now accounting for 55% of complex surgeries in developed nations.
About 58% of global business leaders are currently using physical AI to some extent in operations, growing to 80% who plan to within the next two years.
Where physical AI still struggles
Honesty matters here. Most humanoid platforms still struggle with an "operational ceiling" of three to four hours of battery life, and dexterity for delicate tasks — like threading a needle or handling fragile items — still lags behind human capability.
Despite achieving 95% accuracy in controlled environments, performance can drop to 60% in actual conditions due to environmental factors.
Physical AI is real and accelerating, but the gap between lab performance and field performance remains significant. The honest picture: physical AI is transforming warehouses, factories, and surgical suites now. It's coming to service industries and homes in the next five years. But the vision of a fully capable general-purpose robot doing everything a human can do remains a decade or more away.
What physical AI means for Korea
This is particularly relevant for readers building in the Korean technology ecosystem.
Korea is among the world's leaders in robotics manufacturing, semiconductor technology, and smart factory deployment. South Korea-based startup RLWRLD is developing foundational AI models that allow traditional manual-intensive processes to be performed autonomously by robots through automated learning and mimicking human expertise.
The Korean government's smart city and smart factory initiatives — many of which are part of the very ecosystem this series was written alongside — are a direct entry point into the physical AI wave. The K-startup ecosystem has an unusual opportunity: deep hardware manufacturing expertise, government support for smart infrastructure, and proximity to the Asian markets where physical AI deployment will scale fastest.
Wave 3 on the horizon: AI that improves itself
The third wave is the most speculative — but also the most consequential if it arrives as some believe it will.
Current AI systems are trained, deployed, and then static. Claude Opus 4.6 doesn't get smarter from your conversations with it. GPT-5.2 learns nothing from its interactions in deployment. The model you use today is the same model you'll use in six months — unless Anthropic or OpenAI releases an update.
The next frontier: AI systems that improve themselves through interaction — that learn from each deployment, each failure, each new piece of information — and get measurably better over time without retraining from scratch.
This isn't happening at scale yet. But the research directions are clear:
Continual learning — systems that update their weights from new data without catastrophic forgetting of what they already knew.
Self-play and self-improvement — systems that generate their own training signal by competing against previous versions of themselves (the approach that made AlphaGo superhuman at chess applies to reasoning too).
Recursive improvement — systems that help improve their own training process, generating better data, better evaluation criteria, better fine-tuning approaches.
If this wave arrives, it changes the fundamental dynamic of the field. Right now, improvements come in discrete jumps — model releases every six to twelve months, each significantly better than the last. Self-improving systems would make progress continuous rather than discrete — always getting better, always adapting.
The implications for builders: the systems you build on top of would improve without you doing anything. But the systems you compete with would improve without anyone doing anything either. The moat would have to come from your data, your domain expertise, your relationships — not the underlying model capability, which everyone would have access to at roughly the same rate.
What this means for builders today
Three concrete implications from everything above:
1. Your domain knowledge is the moat
As the tools get better at the generic parts — writing code, generating content, reasoning about problems — the value shifts to what only you know. Your industry expertise. Your customer relationships. Your understanding of a specific domain's nuances. Your proprietary data.
A generic AI system can help anyone write a marketing email. A system trained on your 20 years of enterprise data engineering experience and 500 students of vibe coding curriculum — that's not generic. The builders who win will be the ones who combine powerful generic AI with deep domain specificity.
2. Build for adaptability, not just for now
The tools will change. A workflow you build today on Claude Opus 4.6 will be available to run on whatever model is 10x more capable in two years. Design your systems so that "swap the model" is easy. Don't hard-code prompts with model-specific quirks. Don't build workflows that only work with one provider's APIs.
The abstraction layer — the logic, the data, the workflow design — is yours. The model underneath is a commodity that gets better over time. Treat it that way.
3. The best time to build your AI fluency is now
Here is the honest truth about the next wave: it will be significantly more powerful than what we have today. And it will make people who already understand the fundamentals — problem definition, data quality, prompting, agents, production deployment, responsible building — dramatically more effective than people starting from zero.
The gap between "understands AI fundamentals" and "doesn't" will widen as the tools get more powerful, not close. More powerful tools in the hands of someone who doesn't understand how to direct them just produces faster mediocrity. More powerful tools in the hands of someone who does understand them compounds their advantage.
You've read this series. You're not starting from zero anymore.
A word on what this doesn't mean
The next wave will generate a lot of fear alongside the excitement. Jobs displaced. Power concentrated. Systems beyond human oversight.
Some of those concerns are legitimate and deserve serious attention — this is not the place to dismiss them. The post on responsible AI (Post 11) covers the builder's obligations in this environment. The regulatory conversation, the alignment research, the labor economics — these are real discussions happening in real places, and they matter.
But for a builder reading this: the technology itself is not the determinant of whether the wave is good or bad. How it's deployed is. Who it serves is. What safeguards are built in is. Those are decisions made by builders — by you.
The builders who think carefully about these questions, who build with the responsibility framework from Post 11 alongside the technical capability from the rest of this series, are the ones who create things worth creating.
Build that way.
The full picture
Let's close by naming where we started and where we've ended up.
We started with frustration — AI not delivering on its promise, the demo gap, the comparison spiral, feeling left behind.
We've covered:
- The mental model shift (Post 1)
- What AI actually can and can't do (Post 2)
- How the stack fits together (Post 3)
- Why data is the real bottleneck (Post 4)
- Why problem definition is the real skill (Post 5)
- How to prompt with precision (Post 6)
- How to build and use agents (Post 7)
- How to vibe code a real product (Post 8)
- How to go from demo to production (Post 10)
- How to build responsibly (Post 11)
- And now: where the next wave is taking us (Post 12)
The thread running through all of it: this is learnable. The tools keep improving. The fundamentals stay the same. The people who get consistently good results from AI — now and in the future — are the ones who think clearly about problems, understand the tools they're using, and build with care.
That's you. Go build.
The complete series
If you found this series valuable, the AI Development Guide by Jaehee Song is the book it draws from — going deeper on every topic covered here, with practical examples, architectures, and frameworks you won't find scattered across YouTube tutorials and blog posts.
This is the only book that covers the full arc: from AI fundamentals to production deployment, from solo builders to enterprise workflows, from today's tools to where the field is heading.
📱 Apple Books ▶️ Google Play Books 🌐 All Platforms (Books2Read)
Thank you for reading the "Build with AI" series. If these posts have been useful, share them with someone who's frustrated with AI and doesn't know why yet.