[Repost] One Agent Might Be Better Than Ten

Why I’m Reposting This

This is the edited show notes from Episode 43 of INDIGO TALK, originally published at indigox.me. The full video is on YouTube (1h12m52s).

Who Are Indigo and Xiaoxiao Li?

Indigo (Jedi Lu) is the creator of “Digital Mirror” and an independent investor with a long-term focus on the intersection of AI, Crypto, and superintelligence. His INDIGO TALK series is known for high-information-density deep conversations.

Professor Xiaoxiao Li leads the Trusted and Efficient AI (TEA) Lab at UBC (University of British Columbia). His academic background spans Yale, Princeton, UBC, and the Vector Institute. His journey from neuroscience into AI systems research has given him a unique dual perspective: using the limitations of human cognition to understand AI’s bottlenecks, and using solutions to human limitations to improve AI’s capabilities.

Why This Is Worth Your Attention

The unique value of this conversation lies in the intersection of academic rigor and practical insight:

1. 30 years of evolution in the definition of Agent — tracing from the 1995 textbook definition to Anthropic’s distinction between Workflow and Agent
2. AI also has a “working memory” ceiling — a single Agent’s performance saturates at around 20–30 Skills, a striking parallel to Hick’s Law in human cognition
3. Single Agent plus skills > multi-Agent collaboration (for sequential tasks) — communication overhead and error accumulation are the hidden tax of multi-Agent systems
4. Swarm Intelligence as an alternative to hierarchical management — AI doesn’t need to mimic human organizations; flocking birds and ant colonies might be better models
5. The sobering 90-90 Rule — we may have just completed the “easy” 90% of AI development

TLDR

• Agent ≠ Chatbot: A true Agent requires the ability to autonomously Plan, execute, self-reflect, and self-correct — a definition unchanged since 1995
• AI has a cognitive saturation point: A single Agent performs best at 20–30 Skills; adding more causes performance to plateau and then decline — the context window is AI’s “working memory”
• Single Agent often outperforms multi-Agent: In sequential tasks, the monolithic efficiency of skipping communication overhead is higher, though role separation and parallel sharding still require multi-Agent architectures
• AI doesn’t need to mimic human organizations: Swarm Learning (the flocking birds / ant colony model) may be better suited to AI’s fundamental nature than hierarchical management
• The exoskeleton metaphor: Over-reliance on AI leads to cognitive atrophy; education needs to shift from “teaching skills” to “teaching meaning and purpose”
• Efficiency vs. expansion: AI’s ultimate value is not making the horse-drawn carriage run faster, but enabling humanity to “take flight” — expanding into entirely new dimensions
• The 90-90 Rule: The last 10% of AI development may be harder than the first 90%; the history of autonomous driving is the best cautionary tale

Article Body

Source: INDIGO TALK EP43 — One Agent Might Be Better Than Ten — Indigo’s Digital Mirror Published: February 23, 2026 Guests: Xiaoxiao Li (UBC Professor / TEA Lab Director), Indigo (Digital Mirror creator / independent investor) Video: YouTube (1h12m52s) | Spotify | Xiaoyuzhou

This episode of INDIGO TALK features Professor Xiaoxiao Li from UBC’s Trusted and Efficient AI (TEA) Lab for a face-to-face conversation in Vancouver. Xiaoxiao’s research background spans Yale, Princeton, UBC, and the Vector Institute. His transition from neuroscience to AI systems research has given him a unique dual perspective: using the limitations of human cognition to understand AI’s bottlenecks, and using solutions to human limitations to improve AI’s capabilities.

This conversation covers the definition and evolution of Agent, the efficiency debate between single-agent and multi-agent systems, AI safety and failure tolerance, concerns about cognitive atrophy, and the fundamental transformation of education in the AI era — it is extremely information-dense, and highly recommended.

Timestamps

• 00:01:34 From Yale to UBC: 12 Years of an AI Researcher
• 00:08:25 What Is an Agent, Really: From the 1995 Textbook to Anthropic’s Definition
• 00:15:00 The Limits of Human Cognition and AI’s Parallel Predicament
• 00:23:00 Silicon-Based vs. Carbon-Based: The Fundamental Differences Between Two Forms of Intelligence
• 00:26:00 Single Agent Plus Skills vs. Multi-Agent Collaboration: The Efficiency Debate
• 00:34:00 AI’s Organizational Form: No Need to Imitate Humans
• 00:46:00 AI Failure Tolerance and Safety Design
• 00:53:00 The Exoskeleton Metaphor: When AI Causes the Brain to “Atrophy”
• 01:02:00 Efficiency vs. Expansion: Two Dimensions of the AI Revolution
• 01:10:11 The Sobering Wake-Up Call of the 90-90 Rule

When we casually toss around the word “Agent,” have we stopped to think that this concept was already defined in a 1995 textbook? When we rush to give AI more and more skills, have we noticed that the human brain collapses instantly when faced with too many options — and that AI may not escape a similar fate?

01 From Yale to UBC: 12 Years of an AI Researcher

Canada’s academic contributions to this wave of the AI revolution are severely underestimated. From Hinton to Sutton, from deep learning to reinforcement learning, many of the sparks of research were lit in Canadian universities.

Professor Xiaoxiao’s academic trajectory is highly representative. His doctoral research at Yale focused on the intersection of neuroscience, neuroimaging, and artificial intelligence. After graduating in 2020, he went to Princeton University for a postdoc, shifting his research direction toward machine learning systems, distributed computing, and AI safety and privacy. In 2021, he joined UBC to lead the Trusted and Efficient AI Lab — the TEA Lab. Last year he was promoted to Associate Professor, and he is also a researcher at the Vector Institute founded by Geoffrey Hinton, with visiting research stints at Google during his sabbaticals. By his count, he has been doing AI research for a full 12 years.

The name TEA Lab itself reveals its core mission: no matter how the underlying AI technology evolves, improving AI’s reliability and efficiency is always a topic worth deep cultivation. In the pre-GPT era, the team focused on the efficiency and security of convolutional neural networks in tasks like image classification and language translation. In the era of large models, these problems have not disappeared — if anything, they have become more urgent and complex.

Indigo specifically highlighted a fact that is often overlooked: the academic foundations of this current AI explosion are, to a large extent, rooted in Canada. Professor Hinton at the University of Toronto is the godfather of deep learning; the University of Alberta has Richard S. Sutton, who wrote The Bitter Lesson and is closely collaborating with Google’s Gemini team; and McGill University in Montreal has produced a huge amount of AI talent. As Indigo put it — Canada does the research, America does the entrepreneurship and makes the money. Professor Xiaoxiao laughed and responded that the AI researchers at UBC are relatively young, but hopefully, given time, some heavyweights will emerge there too.

02 What Is an Agent, Really: From the 1995 Textbook to Anthropic’s Definition

Agent is not a new concept — its academic definition was already clearly established 30 years ago. The core that truly distinguishes a Chatbot, a Workflow, and an Agent is the capacity for autonomous Planning.

In 2025, Agent became the hottest word in tech. From Manus to various open-source products, seemingly overnight every AI application had slapped the Agent label on itself. But Professor Xiaoxiao reminds us that from an academic perspective, Agent has long been anything but a new concept.

He cited the definition from the classic 1995 textbook Artificial Intelligence: A Modern Approach by Peter Norvig and Stuart Russell: an Agent is a system that can perceive its environment, make autonomous decisions based on that environment, and maximize its goals. This definition is consistent with what Richard Sutton later described in the Alberta Plan — “perceiving the environment, self-evolving.” Before large models appeared, this was the consensus academic definition of Agent.

However, the market’s current understanding of Agent is far broader. Some simple conversational pipelines or Workflows have also been labeled as Agents. Professor Xiaoxiao cited a very clear distinction Anthropic makes between Workflow and Agent: a Workflow is a system in which large models and tools interact along a predetermined path; an Agent, by contrast, possesses the ability to autonomously Plan and evolve — it can not only use tools, but even create them.

This distinction is critical. Indigo, speaking from personal experience, was direct in saying that for Planning capability, Anthropic’s Claude is the best he has used. Claude doesn’t just understand what you’re saying — it grasps your intent, forms an appropriate plan, executes efficiently, and when it encounters problems, it self-reflects and revises the plan. In Indigo’s words, it already has “a little metacognition” — it knows what it’s doing. Professor Xiaoxiao has observed the same phenomenon, and he hypothesizes it may be because Anthropic has been building foundational Agent infrastructure from the very beginning. Of course, other models have their own strengths: Gemini is extremely fast, OpenAI’s Codex has significantly improved coding capability, and Qwen, Kimi, MiniMax, and others in China are also rapidly catching up.

But the core question remains unchanged: what truly makes an Agent an Agent is the capacity for autonomous Planning — the ability to independently decompose tasks, allocate resources, execute, and self-correct in service of the user’s goals.

03 The Limits of Human Cognition and AI’s Parallel Predicament

Human cognition has clear limits — limited working memory capacity, limited attention, limited skill capacity. While AI is more powerful than humans in many respects, it displays similar bottlenecks. By studying solutions to human cognitive limitations, we can in turn improve AI’s capabilities.

This was the most unexpected segment of the entire conversation. Professor Xiaoxiao started from cognitive psychology and proposed a profound analogical framework: the limitations of human cognition may help us understand why AI also “makes stupid mistakes.”

He mentioned a classic cognitive science theory — human cognitive capacity is limited. For complex systems in the natural world, the rational capacity possessed by the human brain is far from sufficient to solve problems that require exhaustive enumeration of all possibilities. The most intuitive example is chess: in theory an optimal solution exists, but humans can typically only think 3 to 5 moves ahead, and cannot compute the final win probability at every step.

Even more interesting is Hick’s Law. Researchers give subjects a grid of lights and ask them to press the corresponding button when a light turns on. As the number of lights increases, the subject’s reaction time grows sharply and accuracy drops significantly. This reveals a core physical limitation of the human brain: working memory capacity is finite, and when a task exceeds this capacity, the entire cognitive system collapses instantly.

What about AI? Indigo’s intuitive reaction was: AI shouldn’t have this problem — just add more compute, spin up 1,000 instances and run them in parallel. Professor Xiaoxiao gave a nuanced “yes and no” answer. His research found that when you equip an AI Agent with more and more Skills, its performance doesn’t improve indefinitely — it reaches peak performance at around 20 to 30 Skills; beyond that, performance saturates and even declines.

This finding suggests that while AI has surpassed human cognitive limitations in many dimensions (the multi-head attention mechanism in Transformer, for example, can attend to all information simultaneously), it also has its own physical bottleneck — the context window is its “working memory,” and when the Context becomes very long, AI also gets “Lost in the Middle.”

Indigo made a brilliant analogy here: human attention is limited, and can only focus on a few things at once; while Transformer uses multi-head attention to attend to everything simultaneously, the context is its physical constraint — just as memory always has a physical ceiling.

04 Silicon-Based vs. Carbon-Based: The Fundamental Differences Between Two Forms of Intelligence

Although we habitually use human cognition as an analogy for AI, silicon-based and carbon-based intelligence are fundamentally different at the underlying level — AI can be designed, rapidly replicated, and can break through individual limits through brute-force stacking of compute.

Professor Xiaoxiao made an important correction here. While using human cognition to understand AI is a useful research method, he emphasized that this analogy should not be pushed to extremes — silicon-based and carbon-based intelligence are simply different at the fundamental level.

The human brain uses neural electrical signals to transmit information, shaped over millions of years of natural selection. The core of AI is a generative network under a Transformer architecture that simulates intelligence through mathematical operations. The human brain far surpasses AI in energy efficiency — you see two examples and understand a concept, while AI requires training on massive amounts of data. Humans also have stronger generalized learning ability; this “learning through immersion” style is something AI is still catching up to.

But silicon-based intelligence has some advantages that carbon-based intelligence simply does not possess. First, it can be designed — humans can actively attach external capabilities to AI, turning it into an extremely powerful system, whereas humans currently cannot be designed in this way. Second, it can be rapidly replicated — given sufficient compute and energy, you can instantly produce tens of thousands of perfectly identical agents. Third, it can self-evolve — self-evolving algorithms give AI the ability to iterate autonomously.

Indigo made a vivid summary: on Earth, a new intelligent species has been born, and it evolves at a speed far surpassing carbon-based life. While its individual energy efficiency is lower than the human brain, silicon-based intelligence can be “plugged in” indefinitely, producing a completely different form of intelligence through brute-force compute stacking and rapid replication. This fact may sound a little unsettling, but it is indeed the new reality we must face.

05 Single Agent Plus Skills vs. Multi-Agent Collaboration: The Efficiency Debate

In sequential tasks, a single Agent equipped with appropriate Skills is often more efficient than multi-Agent collaboration, because communication overhead and error accumulation are the primary bottlenecks of multi-Agent systems. However, role separation and parallel processing in complex tasks still require multi-Agent architectures.

This is one of Professor Xiaoxiao’s most noteworthy recent research papers: “When Single-Agent with Skills Replace Multi-Agent Systems and When They Fail.” His team found that for certain types of tasks — especially those that can be handled sequentially — giving one Agent multiple skills is often faster and better than having multiple Agents collaborate.

Why? The reason is essentially the same problem human organizations face. When you cooperate with others to accomplish something, communication itself produces information loss. If one person can accomplish something within their own capabilities, eliminating the communication step is actually more efficient. Professor Xiaoxiao uses cooking as an analogy: if one person can both chop vegetables and cook them, that single person making a dish may be faster than two people working in an assembly line, because when operating alone the context is continuous, and they can flexibly optimize their own workflow.

Multi-Agent collaboration also incurs two additional costs. The first is Token cost — every communication between Agents consumes compute and Tokens. The second is error accumulation — in a chained architecture, as long as one Agent makes an error, that error propagates to all downstream Agents.

But this does not mean multi-Agent architectures have no value. Professor Xiaoxiao carefully noted that his research focuses on specific types of sequential tasks and does not cover all scenarios. In two situations, multi-Agent collaboration remains necessary.

The first is when the task itself requires role separation. For example, in Anthropic’s recently released Agent Team demo, 16 Agents worked together to complete a C compiler, which they then used to recompile the Linux kernel. In complex software engineering tasks like this, the person writing code and the person reviewing code must be separate — just as one person cannot be both the referee and the player. Because a single Agent’s memory space cannot be isolated, it cannot be both producer and reviewer within the same context.

The second is when a task can be effectively parallelized and sharded. Indigo shared his experience using Claude Code: when he gives the AI a coding/transcoding task, the AI itself judges whether to spin up Sub-Agents, breaking the task into five pieces and having five Agents process them simultaneously. More cleverly, the AI dynamically decides when to subcontract and when to handle things itself — just like an outstanding human dispatch coordinator.

06 AI’s Organizational Form: No Need to Imitate Humans

Current multi-Agent architectures have over-imitated human organizational hierarchies, but AI may develop completely different collaboration modes — like flocks of birds or ant colonies, operating efficiently without a Leader.

This section of the discussion was particularly illuminating. Indigo opened with an observation: human organization is an extremely redundant and inefficient system, because humans need to rest, have emotions, and have egos — all of which create enormous friction in organizational collaboration. This is why human society developed the tiered management structure of senior, middle, and executive layers.

But for AI, this logic may be completely inapplicable. Professor Xiaoxiao offered a more forward-looking perspective: he is skeptical of current AI Agent systems that over-imitate human organizational hierarchies. The reasoning pathways of AI may be completely different from those of humans, and its organizational methods need not replicate human hierarchical divisions.

He proposed a particularly vivid alternative model: Swarm Learning — just like the way birds flock. A flock of birds has no Leader, and individual birds have very limited intelligence, yet when a group of birds fly together, they form an extraordinarily efficient organizational form. Ants are the same: individual ants are not smart, and there is no commander, but through simple functional division (carrying food, feeding the queen, guarding) and some kind of emergent coordination, ant colonies display remarkable collective intelligence.

Perhaps future multi-Agent systems won’t need a central dispatcher, but will instead spontaneously form efficient organizations through some kind of decentralized collaboration mechanism. Even more interesting is that Professor Xiaoxiao believes this AI-native form of organization could in turn inspire humans — we may be able to learn from AI’s organizational architecture how to make human teams more efficient.

07 AI Failure Tolerance and Safety Design

Different scenarios have vastly different tolerances for AI errors. Medicine and law require extremely high accuracy, and AI output design — especially Confidence Score and highlighting mechanisms — can be a key means of safety assurance.

Indigo and Professor Xiaoxiao went deep on a practical question: to what degree should we trust AI’s output?

Professor Xiaoxiao described a design approach in the medical field that is well worth borrowing by other industries. When applying AI to medical diagnostic assistance, the system automatically highlights keywords where AI is likely to make errors, linking them to relevant medical literature and Clinical Guidelines, and requiring doctors to verify before proceeding to the next stage of analysis. This is not simply telling users “AI might be wrong” — it embeds a verification mechanism at the product design level.

The core concept is the Confidence Score. AI is fundamentally a probabilistic model, and every output it produces has a corresponding confidence level. If the confidence for a particular word or judgment is only 0.2 (out of 1.0), it should be highlighted, alerting the user that human review is needed. This style of “AI output design” thinking is not limited to medicine — it has important value in any professional field requiring high accuracy, such as scientific research and law.

08 The Exoskeleton Metaphor: When AI Causes the Brain to “Atrophy”

Over-reliance on AI will lead to atrophy of human cognitive abilities, just as wearing an exoskeleton for a long time causes muscles to atrophy. The fundamental mission of education needs to shift from “teaching skills” to “teaching meaning and purpose.”

Anthropic recently published a research report showing that developers who use Claude Code long-term have experienced varying degrees of regression in their programming ability. Professor Xiaoxiao agreed strongly, viewing this as a very serious problem — over-reliance on AI makes the brain lazy.

Indigo offered a particularly apt metaphor: the exoskeleton. Imagine you want to build muscle by lifting weights, but now there’s a high-tech exoskeleton — put it on and lifting becomes effortless. The problem is, you’ve completed the lifting motion without exercising a single muscle. Over time, your muscles will inevitably atrophy. He also used himself as an example: after using Tesla’s FSD autopilot, he noticed his driving skills had clearly deteriorated, so he insists on driving himself one day out of every seven — specifically to maintain basic driving ability.

This exoskeleton metaphor extends to questions about education. Indigo posed a sharp follow-up question: when AI can help you accomplish almost every skill-based task, you sit in front of a computer, facing the AI, but don’t know what to ask — because you have no goal.

The history of education reveals the root of the problem. Since the industrialization era over 200 years ago, the core of education has always been “teaching skills” — training soldiers to follow commands and march in step, training workers to operate machines, training white-collar employees to use software. The goal was clear: learn a skill, then go to work. But when AI can provide these Skills on demand, traditional skill-based education loses its meaning.

Professor Xiaoxiao and Indigo agree that education needs a fundamental shift in direction: from “teaching skills” to “teaching meaning.” Traditional education has never taught you how to seek meaning or formulate your own goals. As more and more “rote work” is replaced by AI, and as the bosses who used to assign you goals no longer need you, you must be capable of forming your own goals.

University education still has meaning, but its value is no longer in transmitting knowledge (AI can already do that) — it lies in cultivating two core thinking abilities: Critical Thinking (the ability to scrutinize problems and answers) and logical reasoning. The disciplines we once considered “useless” — philosophy, theoretical physics, art — may in the AI era actually become more important, because what they cultivate is precisely the kind of general cognitive frameworks and cross-domain creativity that matters most.

09 Efficiency vs. Expansion: Two Dimensions of the AI Revolution

AI is currently primarily improving human efficiency, but its deeper significance may lie in expanding human boundaries. Only by expanding boundaries can the economic “pie” grow larger, rather than everyone simply fighting over the same slice.

Near the end of the conversation, Professor Xiaoxiao proposed a framework that opened new horizons: AI development has two directions — improving efficiency and expanding boundaries.

Improving efficiency is obvious: AI can do what humans already do, faster and better. But this path has a problem — if the total size of the economy (the pie) doesn’t grow, the more efficient simply consume the less efficient, and the end result is a zero-sum game.

Expanding boundaries is what actually makes the pie grow. Professor Xiaoxiao uses the airplane as an analogy: the airplane didn’t make the horse-drawn carriage faster — it let humans fly — and that is a dimension expansion of an entirely new kind. Every invention in history that qualitatively transformed the efficiency of human movement through space (the steam engine, the railroad, the airplane) brought entirely new industries and occupations. And this wave of the AI revolution is like what Jobs once said about computers being “a bicycle for the mind” — AI is a true intellectual accelerator, and it may bring human cognition and creativity into an entirely new dimension.

Indigo extended this logic further: if AI frees up large amounts of repetitive labor, and people have more time to invest in creation and exploration, we may be entering a new Renaissance. Of course, there is a very realistic transition period in between — people must first meet their basic needs before pursuing meaning.

10 The Sobering Wake-Up Call of the 90-90 Rule

The 90-90 Rule from software engineering applies equally to AI development — 90% of functionality took 90% of the time to complete, but the remaining 10% also requires 90% of the time. The current optimism about AI progress may be premature.

At the very end of the conversation, Professor Xiaoxiao contributed a beautifully sobering conclusion. He cited the famous Ninety-Ninety Rule from software engineering: 90% of the code takes 90% of the time to complete, but the remaining 10% of the code also takes 90% of the time. This means we may have only completed the relatively “easy” part of AI development, and the final 10% — or even 0.1% — may be far harder than anyone imagines.

He used the history of autonomous vehicle development to support this point — a few years ago everyone thought fully autonomous driving was just around the corner, yet the leap from L2 to L5 has been far more difficult than expected. Other areas of AI will very likely follow the same pattern.

This clear-eyed reminder provided a perfectly fitting conclusion to a conversation that had been full of excitement and imagination. The future is both good and not good — this ambiguous state is precisely the objective reality.

When Indigo excitedly describes AI’s scheduling capabilities, Professor Xiaoxiao uses Swarm Learning to remind us: don’t assume AI’s organizational form must imitate humans. When Professor Xiaoxiao uses the framework of cognitive psychology to analyze AI’s limitations, Indigo points out that silicon-based intelligence can overcome many physical constraints humans face through brute-force compute stacking.

The deeper reflection is this: if AI can truly expand humanity’s boundaries rather than merely improve efficiency, then the significance of this revolution far exceeds the short-term questions we’re currently preoccupied with — “who will lose their job,” “which model is better.” What education needs to teach is no longer skills, but a sense of purpose and meaning. This sounds idealistic, but at a moment when AI is evolving at an exponential pace, it may be the most pragmatic advice of all.

Key Takeaways

On the nature of Agent: An Agent is not the same as a Chatbot, nor is it the same as a Workflow. A true Agent must possess the ability to autonomously Plan, execute, self-reflect, and self-correct. This definition has not fundamentally changed from 1995 to today.

On cognitive limitations: The human brain’s working memory capacity is limited; when tasks exceed this capacity, performance collapses (Hick’s Law). AI has a similar saturation point — a single Agent performs optimally at 20–30 Skills, after which performance plateaus.

On single Agent vs. multi-Agent: In sequential tasks, a single Agent with skills is often more efficient than multi-Agent collaboration, because communication overhead and error accumulation are multi-Agent’s two major bottlenecks. However, for complex tasks requiring role separation and parallel processing, multi-Agent architecture is irreplaceable.

On AI’s organizational form: There is no need to have AI replicate human hierarchical management models. Decentralized collaboration like Swarm Intelligence may be better suited to AI’s fundamental nature — and may even in turn inspire transformations in human organizational design.

On cognitive atrophy: Over-reliance on AI will lead to cognitive regression, just as an exoskeleton will cause muscles to atrophy. Preserving the capacity for independent thinking requires deliberate practice — just as people go to the gym after physical labor becomes automated.

On educational transformation: The core of traditional education is “teaching skills,” while the education of the AI era needs to shift toward “teaching meaning and purpose.” Critical Thinking, logical reasoning, and cross-domain creativity — these will become humans’ most irreplaceable capabilities.

On staying clear-headed: The 90-90 Rule reminds us that the last 10% of AI development may be harder than the first 90%. The history of autonomous driving is the best cautionary tale.