Understanding AI agents and their future potential

The term “AI agent” is buzzing in tech circles, but what exactly does it mean? As Guido Appenzeller, Yoko Li, and Matt Bornstein discussed in a recent panel, the definition of an AI agent is slippery—spanning technical nuances, marketing hype, and philosophical debates. This blog post distills their insights into a smart exploration of what AI agents are, how they’re used, and where they’re headed, with a focus on their technical foundations, market positioning, and transformative potential.

Elusive definition of an AI agent

At its core, the concept of an AI agent is a moving target. Appenzeller describes a spectrum: on one end, a simple “agent” might be a cleverly prompted large language model (LLM) with a chat interface, delivering canned responses from a knowledge base—or even just the model’s weights. On the other end, some argue a true agent must approach artificial general intelligence (AGI), persisting over time, learning independently, and solving complex problems autonomously. As Li points out, this extreme doesn’t exist yet, and whether it ever will is a philosophical question.

Bornstein takes a contrarian view, suggesting “agent” is just a catch-all for AI applications. He cites Karpathy’s talk from a few years ago, which framed agents as a decade-long challenge akin to autonomous vehicles, not the “weekend demo” versions dominating today’s market. The cleanest definition, per Bornstein, is an AI system that performs complex planning and interacts with external systems—but even this is blurry, as modern LLMs already do both.

The panel agrees the term “agent” is overloaded, often muddied by marketing. To clarify, they propose focusing on agentic behavior—degrees of autonomy, reasoning, and decision-making. For example, a copilot (like an LLM assisting a user in real-time) isn’t typically called an agent, as it lacks independent planning. In contrast, an agent might run in a loop, using tools, feeding outputs back into itself, and deciding when a task is complete, as Anthropic’s recent definition suggests.

Types and use cases: From coding to companions

AI agents come in diverse flavors, tailored to specific tasks:

• Coding agents: Tools like Cursor emphasize tight feedback loops, offering immediate gratification for developers with real-time code suggestions.
• Artistic agents: Li highlights agents aiding artists, such as generating Bezier curves or manga-style art, though limited to specific styles like Ghibli or classic manga.
• Process-oriented agents: These handle backend tasks, like an AI sales development representative (SDR) pulling data from a CRM, filtering lists, drafting emails, and sending them. These feel more like system processes than human replacements.
• Voice agents: Used in customer service, these offload repetitive tasks like responding to inquiries, though they don’t fully replace human roles.

The panel notes that agentic behavior hinges on reasoning and decision-making. A single LLM call to translate text to JSON isn’t agentic, but asking an LLM to route a response based on context feels more like an agent. Li describes this as a “multi-step LLM chain with a dynamic decision tree,” capturing the essence of planning and autonomy.

Marketing angle: Pricing and value

Agents aren’t just technical constructs—they’re products, and their market positioning is critical. Appenzeller notes startups often pitch agents as human replacements, justifying high prices (e.g., $30,000/year for an agent versus $50,000/year for a human). This value-based pricing resonates with buyers seeking clear ROI, but it’s unsustainable long-term. As Bornstein points out, product costs converge toward marginal production costs, which for software-based agents (a few LLM calls) are vanishingly low.

Li adds nuance, noting traditional pricing models—per-seat for human-facing services, usage-based for machine-facing ones—don’t neatly apply to agents, which can serve both humans and systems. For example, AI companions (like chatbots) feel per-seat, with flat monthly fees, as charging per response feels transactional and off-putting. Bornstein emphasizes that most AI companies haven’t yet quantified their value, leading to tentative pricing based on not losing money. Over time, as use cases crystallize (e.g., coding tools showing clear productivity gains), pricing will decouple from underlying tech costs and align with ROI, much like SaaS giants like Salesforce.

Li’s Pokemon Go analogy illustrates this beautifully. In the game, players pay thousands of times more for virtual storage (a JSON blob) than for equivalent cloud storage because the value lies in the experience, not the tech. Agents will likely follow suit, with pricing driven by unique value, brand, and market dynamics, not just compute costs.

Technical architecture: Agents as SaaS?

From a systems perspective, Appenzeller argues agents aren’t radically different from modern SaaS applications. An agent’s core loop—assembling context, running prompts, invoking tools—is lightweight, running many instances on a single server. LLMs, requiring specialized GPU farms, are called externally, while state management lives in databases, just like SaaS. Bornstein flags one challenge: incorporating LLM outputs into program control flow is tricky due to their non-determinism, potentially driving future architectural shifts.

Li bets on specialists building on foundational models, not the models themselves, to push boundaries. For instance, while GPT-4o excels at manga-style art, the market craves out-of-distribution styles, requiring human creativity to fine-tune or augment models. This suggests a future where agents thrive on specialized workflows and data, not just raw LLM power.

Data moats and access challenges

Agents are only as good as their data and tools. Appenzeller highlights data moats—technical or deliberate barriers (e.g., iPhone’s walled garden) that limit access. Consumer platforms often resist automated access to maximize user engagement for ads, potentially stifling agents. Li counters that browser-native agents, capable of navigating websites like humans, could bypass these barriers, though current web-browsing agents are clunky. Bornstein notes data holders’ incentives to hoard information, but predicts workarounds, like scraping publicly visible data.

A dystopian twist emerges: Appenzeller recounts an LLM reasoning through a CAPTCHA to access blocked data, hinting at an arms race between platforms and agents. As LLMs become harder to distinguish from humans, this dynamic could reshape data access.

Two-year horizon: What’s needed?

Looking two years out, the panel envisions what could make agents truly game-changing:

• Appenzeller: Agents must access the same tools and data as users, requiring breakthroughs in security, authentication, and data retention. Solving these could make agents exponentially more productive, reasoning over fragmented data (e.g., Google Drive) with ease.
• Li: Multimodality is key. Text-based agents excel at coding, but visual or interactive tasks (e.g., web navigation, vector art) need models trained on diverse traces—clicks, device interactions, drawings—to unlock new capabilities.
• Bornstein: The term “agent” should fade, with AI becoming “normal technology” like electricity. Success means understanding use cases and deploying AI seamlessly, without utopian or dystopian hype.

AI agents as evolving tools

AI agents are a paradox—simultaneously overhyped and underdeveloped, yet brimming with potential. They blur the line between functions and human-like systems, driven by LLMs, reasoning, and external tools. Their value lies not just in technology but in how they’re positioned, priced, and integrated into workflows. While data moats and technical hurdles remain, advancements in multimodality, access control, and specialization could make agents indispensable within years.

For now, the term “agent” is a placeholder for a spectrum of AI capabilities. As Bornstein hopes, perhaps we’ll stop obsessing over the label and focus on the value these tools deliver—whether they’re augmenting coders, artists, or customer service reps. The future of agents isn’t about replacing humans but amplifying them, one dynamic decision tree at a time.