Investment Blog

Next AI Wave: Beyond Chatbots to Action and Reasoning

If you're asking what the next AI wave is, you're likely tired of the hype. Chatbots are impressive, but they're just the opening act. The real shift, the one that will redefine industries and create new investment landscapes, is moving from AI that talks to AI that acts and reasons. I've spent months testing early agentic systems, talking to researchers at labs pushing the boundaries of reasoning, and the consensus is clear: the next wave is about autonomy and embodied intelligence. It's less about generating a perfect paragraph and more about an AI that can plan a complex project, execute the steps across different software, learn from mistakes, and even interact with the physical world through robots. This isn't a distant dream; the foundational pieces are being assembled now.

From Chat to Action: The Rise of AI Agents

Think of today's popular AI as a brilliant, fast-talking consultant. You ask a question, it gives an answer—sometimes stunning, sometimes hilariously wrong. But it stops there. An AI agent is that consultant who doesn't just give advice; it gets up, makes the calls, books the flights, negotiates the deal, and writes the report, all while adapting to unexpected hiccups.

How does it work? It's a software loop. The agent is given a high-level goal—"plan and book a cost-effective 5-day business trip to Berlin with three client meetings." Instead of spitting out a text itinerary, it breaks this down. It might first call a reasoning module to strategize. Then, it starts executing: opening a browser to search for flights, logging into a corporate travel tool to check policies, accessing your calendar to find free slots, drafting emails to clients to propose times, and finally, booking the hotel that fits the per-diem. It does this by controlling your computer's cursor and keyboard, or through specialized APIs.

I've run tests with several open-source agent frameworks. The promise is intoxicating. You can literally watch the code it writes, the websites it navigates. But here's the non-consensus truth most demos gloss over: current agents are incredibly fragile. They get stuck in loops. They click the wrong button. A pop-up ad can derail the entire operation. The reliability is maybe 60-70% on a good day for a simple task. This fragility is the primary bottleneck, not the core intelligence. Investing now means betting on teams solving this reliability engineering problem, not just making a slightly smarter chatbot.

The Core Shift: We're moving from statistical pattern matching (predicting the next word) to goal-directed planning and execution. The key metric changes from "Is this answer plausible?" to "Did the task get completed successfully and efficiently?"

Where You'll See Agents First (And Where You Won't)

Don't expect a general-purpose agent to run your life next year. The rollout will be vertical and specific.

Early Wins: Customer service triage that actually resolves issues by accessing backend systems. Automated data analysis and report generation where the agent queries databases, cleans the data, creates visualizations, and writes insights. Personal coding assistants that don't just suggest code but debug, run tests, and implement the fix.

Later, Harder Problems: Fully autonomous scientific discovery (though we'll see assistive tools first). Holistic business management. Any task requiring deep, nuanced understanding of an unpredictable physical environment.

The Missing Piece: Building a True Reasoning Engine

Here's a subtle mistake many make: conflating knowledge with reasoning. Current large language models have vast knowledge, but their reasoning is often a shallow mimicry of logic seen in training data. The next wave demands systems that can reason novelly—to tackle a problem they've never seen before by breaking it down using principles, not just by recalling a similar example.

Why is this so critical for agents? Because the real world is full of novel situations. If your agent's plan to book a flight fails because the website is down, it needs to reason: "OK, the primary path is blocked. Can I use a different airline's site? Can I call the travel agency API instead? If not, should I wait and retry, or notify the user immediately?" This requires causal understanding and contingency planning.

The research is moving beyond simply scaling data. Look at projects like OpenAI's reported search for "reasoning neurons" or Google DeepMind's work on chain-of-thought prompting and its successors. The frontier is in architectures that separate planning from action, that have an internal "workspace" to simulate outcomes before acting—a kind of mental rehearsal. Some of the most exciting papers on arXiv right now are about improving mathematical reasoning and planning in abstract environments, which are proxies for this broader capability.

From an investment perspective, this means the companies or open-source projects that crack more efficient, reliable reasoning models will become the indispensable engine powering all agents. It's a layer beneath the application.

The Physical Frontier: When AI Gets a Body

This is the part that feels like science fiction but is accelerating in labs. Embodied AI refers to intelligence that learns by interacting with a physical environment, whether through a robot, a car, or even a virtual simulation. The next wave isn't just digital; it has eyes, hands, and wheels.

The learning is fundamentally different. Instead of learning from trillions of text tokens, an embodied AI learns from millions of physical trials—picking up a mug, navigating a cluttered room, folding a towel. This creates a rich, multi-sensory understanding of physics, cause-and-effect, and space that pure text models lack. Companies like Covariant, Boston Dynamics (with Hyundai), and Tesla (with its Optimus robot) are betting big here.

I spoke with a researcher working on robotic manipulation. He said the biggest challenge isn't the mechanical grip, but the AI's understanding of "material compliance"—how a soft bag deforms versus a rigid box. This is knowledge you can't just read about; you have to experience it through thousands of failed attempts. This wave will first hit structured environments: warehouses, factory floors, surgical suites. The long-term vision—a robot helper at home—remains a massive challenge due to the insane complexity and variability of our living spaces.

How to Position Yourself for the Next AI Wave

If you view this through an investment lens, the landscape looks different from the ChatGPT frenzy. It's less about a single consumer app and more about infrastructure, enabling technologies, and specific vertical applications.

Focus Area What It Means Potential & Risk Profile
Agent Infrastructure & Platforms Tools that help developers build, test, and deploy reliable agents (e.g., orchestration frameworks, evaluation suites, safety layers). High Potential, Moderate Risk. The "picks and shovels" play. Winners could become the standard OS for autonomous AI.
Vertical-Specific Agents Agents built for one industry: legal document review, medical diagnosis support, automated scientific experimentation. Variable. Success depends on deep domain expertise and solving that industry's unique reliability hurdles. High reward if they crack it.
Robotics & Embodied AI Software The AI brains for physical systems, not necessarily the robot hardware itself. Perception, control, and task-planning software. High Risk, Long Timeline. Capital-intensive and slow to mature, but the upside for winners is foundational.
Next-Gen Reasoning Models Companies or research pushing beyond transformer-based next-token prediction to new architectures for planning and logic. Speculative, Asymmetric Upside. Could be a breakthrough that reshapes the field, or a research dead-end.

A personal take most analysts miss: don't sleep on the simulation software companies. Training embodied AI in the real world is slow and expensive. Training it in hyper-realistic virtual worlds is faster and cheaper. The demand for high-fidelity physical and interactive simulations will explode.

The skill set is changing too. The most valuable people will be those who can bridge AI capabilities with real-world constraints—not just ML engineers, but also systems integrators, reliability engineers, and ethicists who can design boundaries for these autonomous systems.

Your Questions on the Future of AI

Are AI agents reliable enough to use for anything important right now?
For most critical tasks, no, not fully autonomously. The failure modes are too unpredictable. Where they excel today is as super-powered assistants under human supervision. You give an agent a sub-task—"compile the sales data from these three sources into one spreadsheet"—and you review its work. The value is in dramatic acceleration, not replacement. Blind trust is a recipe for disaster. The trajectory, however, is toward increasing reliability and scope of autonomy.
Is the next wave the same as Artificial General Intelligence (AGI)?
Not exactly, but it's a major stepping stone. AGI implies human-like adaptability across any intellectual task. The next wave of action-oriented, reasoning AI gets us closer by tackling key AGI components: autonomous goal achievement, planning, and interacting with open environments. Think of it as moving from narrow AI (chat) to broad, but not yet general, AI. It's expanding the range of tasks AI can perform without human micro-management.
What's the biggest barrier to this next wave becoming mainstream?
Beyond technical fragility, it's the "interface problem." Today's world—websites, software, office protocols—is built for human cognition. It's messy, inconsistent, and full of implicit knowledge. For an AI agent to operate seamlessly, either the world needs to be restructured with machine-readable standards (a huge undertaking), or agents need to become incredibly robust at navigating human-designed chaos. The latter is where most of the work is focused, and it's a brutally hard problem.
As an investor, is it too early or too late to look at this space?
It's early for the specific applications of the next wave, but the foundational companies are already being built and funded. The chatbot wave identified the appetite. Now is the time for due diligence on teams building the underlying architecture for action and reasoning. Look for companies with a clear handle on the reliability challenge, not just a flashy demo. The vertical applications will take 3-5 years to mature meaningfully, but the infrastructure bets are happening now.
What's a realistic personal use case I might see in the next 2-3 years?
A truly effective personal administrative agent. One that reliably handles the drudgery: not just drafting emails, but sorting your inbox, reconciling expenses from receipts, scheduling meetings by negotiating times with other people's agents, managing your personal knowledge base, and handling routine online purchases and bookings. It would work across your approved apps and always ask for confirmation before any final action. This solves a real pain point—administrative overload—and operates in a relatively structured digital domain.

The path forward is clear. The next AI wave is leaving the chat window and entering the world of action, reasoning, and physical interaction. It will be messier, harder to build, and more transformative than what we've seen so far. The time to understand its contours isn't when it's already here, but now, as the foundations are being laid.

Next Positive Trends in the Industrial Economy

Leave a comment