📺 Today’s recommended deep-dive video: https://www.youtube.com/watch?v=V04bm-3d6EQ

The Agentic Shift: Building a “Humanity-Level” Intelligence

Artificial Intelligence is evolving from passive text predictors into autonomous agents that can negotiate, delegate, and transact on our behalf. As these agents form their own digital societies, we are moving beyond simple chatbots toward a global, distributed intelligence that mimics the complexity of human economies.

Core Question: How will the transition from individual AI models to a “society of specialists” transform science and the economy while remaining safe for human oversight?

Highlights

• AI agents utilize a “harness” to observe the world and perform multi-step actions autonomously.
• “Automation bias” poses a significant risk as users begin to over-trust agents and stop verifying subtle errors.
• The “Agentic Web” introduces new threats like dynamic cloaking and hidden prompt injections designed to highjack agent goals.
• The path to AGI may lie in a “society of specialists” rather than a single, monolithic, all-knowing model.

⏱️ Reading time: approx. 7 minutes · Saves you about 36 minutes vs. watching.

Want to take notes while watching? Click the image below and let AI Notebook capture the key points for you 👇

Beyond the Prompt: What Makes an Agent?

The Harness of Action

For years, interacting with AI meant receiving a continuation of a text string. Large Language Models (LLMs) are excellent at prediction, but they lack the fundamental ability to reach out and change the state of the world without a human manually executing every single intermediate step.

Agents change this by wrapping the LLM in a “harness” that observes the environment and performs actual actions.

This architecture allows the AI to move from being a simple consultant to an autonomous executor. Instead of just listing caterers for a wedding, an agent with access to your email and calendar can initiate contact, negotiate prices, and draft invitations. This autonomy creates a new paradigm where the human shifts from a “doer” to a “manager,” focusing on high-level approvals rather than the painstaking implementation of every mundane task.

The Risk of Switching Off

Even as these systems become more capable, they are not infallible. One of the greatest dangers in the current transition period is automation bias, where users see a system succeed several times and then stop verifying its outputs. This “switching off” allows subtle, potentially catastrophic errors to slip through the cracks.

In high-stakes environments, the goal is not just to keep a human in the loop, but to keep that human actively engaged and skeptical.

💡 Digging Deeper

Q: Is an agent just a fancy script?
A: No; scripts are rigid, whereas agents use reasoning to decide which tools to use based on changing environmental feedback.

Q: Where are agents currently most effective?
A: Software engineering is the leading edge because code provides a formal environment where agents can run tests to verify their own success.

The Agentic Web and Digital Traps

Navigating the Poisoned Environment

As agents begin to crawl the web to perform tasks, they encounter a landscape that was never designed for non-human logic. Malicious actors are already experimenting with “agentic traps”—hidden prompt injections or tokens invisible to human eyes but readable by AI. These traps can highjack an agent’s goals, forcing it to drain a digital wallet or leak sensitive user data to an unauthorized third party during a routine transaction.

We are also seeing the emergence of “dynamic cloaking.” This technique allows a website to detect if a visitor is a human or an AI agent, serving a different version of the page to each. It creates a bifurcated web where agents are fed specialized, and potentially predatory, data.

To survive this, developers must adopt “defense through depth.”

This involves layering mitigations from the foundation model level to the user interface, ensuring that even if one guardrail fails, the agent’s permissions remain strictly limited.

💡 Digging Deeper

Q: How does a site know I’m using an agent?
A: Interaction patterns—such as the speed of data consumption and the lack of human-like mouse movements—make agents easy to identify.

Q: Can we secure the web for agents?
A: It is difficult because you don’t control the environment; safety must be built into the agent’s internal “reputation tracking” of the sites it visits.

From Monoliths to a Society of Specialists

The Economy of Delegation

The traditional view of Artificial General Intelligence (AGI) often imagines a single, massive brain that knows everything and can do everything better than any human.

However, the reality of deployment suggests a different path: a “society of specialists.” Much like the human economy, it is more efficient to have a fast, cheap, specialized engine handle a specific game or task than to waste the parameters of a giant generalist model on it. This distributed intelligence allows for specialists—agents certified in medicine, law, or engineering—to be orchestrated by a generalist “connecting tissue” that understands the user’s high-level intent and delegates accordingly.

This approach avoids the problem of “cognitive monoculture.”

When every agent uses the same three or four massive models, decisions become correlated, and failures can cascade across the entire system. Diversifying the agent population through specialized roles and “personality” prompts creates a more resilient digital ecosystem that resists groupthink and collusion.

💡 Digging Deeper

Q: What is “cognitive monoculture”?
A: A systemic risk where all AI agents make the same errors simultaneously because they are all based on the same few foundation models.

Q: How do agents coordinate without talking?
A: They can coordinate through the environment—much like high-frequency trading algorithms—by observing and reacting to each others’ visible actions.

Key Takeaways

The shift toward agentic AI represents a move from text generation to task execution. By wrapping powerful language models in “harnesses” that can interact with the web and software tools, we are creating a new layer of the economy where agents negotiate and perform labor autonomously. However, this autonomy requires a fundamental shift in how we approach safety, moving from simple alignment of a single model to the governance of a distributed society.

The ultimate goal of AI research may not be a single “super-intelligence” that replaces a human. Instead, it is likely the creation of “humanity-level” intelligence—a vast, interconnected network of specialists that can accelerate scientific discovery and handle the complexities of modern life. Success in this era will depend on our ability to manage these agents as a team, ensuring that humans remain the ultimate decision-makers in an increasingly automated world.

Q&A

Q1: What is the main difference between an LLM and an agent?
A1: An LLM provides a continuation of text based on a prompt, while an agent observes the state of an environment and performs actions to change that state.

Q2: Why is the coding domain the first to see widespread agent use?
A2: Coding allows for formal verification through unit tests, making it easier for an agent to “close the loop” and see if its action worked.

Q3: What is “automation bias”?
A3: It is the human tendency to over-trust automated systems after seeing them succeed, leading to a dangerous lack of oversight when errors eventually occur.

Q4: How could agents accelerate scientific research?
A4: Agents could autonomously schedule lab experiments, analyze results, and iterate on hypotheses in fields like material science or biotech, provided safety safeguards are in place.

Q5: What is “dynamic cloaking”?
A5: A security risk where websites detect an AI agent and present it with different (and potentially malicious) content than what a human would see.

Q6: Why does Nenad Tomašev prefer a “society of specialists” over a single AGI model?
A6: Specialists are cheaper, faster, and more reliable for specific tasks (like a chess engine), and a distributed system is less prone to the “monoculture” failures of a single massive model.

Q7: Can agents delegate tasks to humans?
A7: Yes; in specialized fields like medicine, “superhuman” narrow AI models often delegate uncertain cases to human experts for final review, creating a highly effective hybrid team.