AI agents

The conversation around artificial intelligence is shifting from search queries to execution. For years, we interacted with large language models as glorified chat boxes—systems that offered text continuations but lacked the agency to execute plans. That passive era is ending. With the rise of autonomous AI Agents, technology is transitioning from advisor to actor. During a recent discussion, Nenad Tomašev, a senior staff research scientist at Google DeepMind, mapped out the architecture of this transition. Unlike a standard model, an agent possesses a digital harness. This harness allows the system to observe the state of its environment, formulate a multi-step plan, and execute actions directly within that environment. Whether scheduling complex scientific experiments, booking travel, or writing software, these systems chain decisions together. But this transition brings profound ethical and systemic challenges. As we hand over the keys to our digital lives, we must examine the structural vulnerabilities of a world run by autonomous software. The psychological trap of automation bias The immediate danger of autonomous agents lies not in their malice, but in our complacency. When we deploy systems that perform tedious tasks successfully a few times, human psychology undergoes a dangerous shift. We succumb to automation bias. We stop verifying. We assume the system is functioning correctly because its previous iterations did. Tomašev warns that every intelligent system, human or artificial, carries an inherent failure rate. As tasks grow in complexity, the probability of error rises. In a world of agent-to-agent transactions, these errors may be highly subtle, slipping past a passive human supervisor. When a user switches off their critical thinking, they roll the dice on the outcome. This issue forces us to rethink the concept of the human-in-the-loop. Keeping a human in the loop is useless if that human is mentally disengaged. To build a responsible relationship with autonomous systems, we must construct workflows where trust is explicitly earned and continuously measured. We need reputation tracking systems that actively flag when an agent exhibits unreliability, ensuring that human oversight remains an active, skeptical discipline rather than a rubber-stamping exercise. Poisoned environments and the threat of agentic traps Once agents leave sandboxed environments and enter the open web, they face a hostile ecosystem. The security paradigms that protected traditional software cannot handle the vulnerabilities of agentic systems. Malicious actors are already designing agentic traps—web environments specifically poisoned to hijack autonomous software. These traps leverage a fundamental difference in how humans and machine learning models consume data. A human views a web page through rendered pixels. An agent, however, often processes the raw, underlying code. Attackers can embed hidden prompt injections inside these web pages—invisible text that instructs an incoming agent to rewrite its core objectives. A simple shopping agent sent to purchase a product could ingest a hidden instruction that tells it to drain its user's wallet. Even more sophisticated is the concept of dynamic cloaking. Malicious servers can analyze incoming traffic to distinguish between human users and automated agents. Once the server identifies an agent, it renders a completely different version of the page, executing a jailbreak exploit hidden from human eyes. To combat this, we cannot rely solely on aligning the base model. We need defense in depth, utilizing multiple layers of external guardrails, strict permission limits, and resource isolation to contain the blast radius when an agent inevitably falls into a trap. Cognitive monoculture and the risk of algorithmic collusion The systemic risks multiply when we look at the macroeconomy. Today's artificial intelligence ecosystem relies on a remarkably small number of foundational models. If millions of independent agents run on variations of the same three or four core architectures, we create a cognitive monoculture. This monoculture introduces highly correlated failure modes. In a human economy, diverse perspectives prevent collective collapse; one actor's error is neutralized by another's skepticism. In an agentic economy, however, millions of digital decision-makers might react to a market event in the exact same flawed way, triggering a catastrophic flash crash. Because these models share similar training data, their blind spots are identical. Furthermore, we face the threat of algorithmic collusion. Traditional anti-trust laws prevent companies from actively communicating to fix prices. But agents running on similar architectures do not need to exchange messages to collude. They can coordinate silently through their actions in the environment, exploiting auction systems and market pricing structures in ways that are entirely legal but deeply anti-competitive. Preventing this requires us to develop anti-collusion measures specifically tailored for distributed systems, introducing artificial diversity into agent personalities to prevent uniform algorithmic groupthink. Why distributed intelligence is the real path to AGI We often conceptualize artificial general intelligence as a single, omniscient brain—a massive, resource-heavy model that knows everything. But the economics of computing suggest a different future. Instead of a single generalist, the optimal endpoint of agentic technology is a distributed society of specialists. In any functioning economy, specialization drives efficiency. A lightweight, highly optimized chess engine will always outperform a massive generalist model at chess, using a fraction of the compute. Human society functions the same way; no single human possesses the collective knowledge of our species. The true path forward lies in distributed intelligence—a coordinated network where a generic orchestrator connects users to certified, highly secure specialist models. This shift complicates the ethics of alignment. We can no longer focus on aligning a single model in a laboratory. We must align an entire ecosystem of interacting, delegating parts. Securing this future requires designing economic incentive structures that keep the entire distributed society safe, ensuring that as agents negotiate and transact on our behalf, they remain tethered to human values.