The agentic AI era
Posted on: 17 December 2025
Google launched Gemini 2.0 Flash with explicit framing: "our new AI model for the agentic era". This isn't a marketing detail. It's a field-defining operation that deserves clinical attention.
CEO Sundar Pichai described agents as systems that "understand more about the world around you, think multiple steps ahead, and take action on your behalf, with your supervision". That final clause, "with your supervision", is the key to understanding the entire paradigm now emerging.
The precedent that illuminates
Every time a tool acquires autonomous action capability, the pattern is predictable. First phase: enthusiasm for possibilities. Second phase: the first large-scale incident revealing a structural flaw. Third phase: retroactive regulation.
We saw it with algorithmic trading. For years, algorithms operated with theoretical human supervision but growing practical autonomy. Then came the flash crash of 6 May 2010: a thousand-point Dow Jones collapse in minutes, caused by algorithms interacting in ways nobody had predicted. Human supervision existed on paper. In reality, systems operated at speeds that made human intervention structurally impossible.
Today's AI agents operate in an analogous regime, but with a critical difference: they're not confined to a specific domain like financial markets. Google's Project Mariner operates in the browser. Jules operates in code. Project Astra aspires to be a "universal assistant". The potential for interconnection is of a different order entirely.
The architecture of responsibility
Why does Google explicitly define this as the "agentic era"? The answer lies in the economics of responsibility.
Whoever defines the paradigm also defines success criteria and, implicitly, the boundaries of blame. "With your supervision" isn't a feature description. It's a risk allocation clause. The agent acts, but final responsibility remains with whoever supervises.
It's the same pattern platforms have used for years: "We provide the tool, you're responsible for how you use it." This worked as long as tools were passive. With agents that "think multiple steps ahead and take action", boundaries become porous. If the agent chains ten actions and the seventh causes damage, should the supervisor have intervened at the sixth? The third? Before starting?
The legal answer will arrive after the first significant incidents. Meanwhile, anyone adopting these systems is implicitly accepting a liability regime that remains undefined.
Two ways to use an agent
There's a real bifurcation in adoption of these systems, but it's not the one that appears on the surface.
It's not "AI spam vs intelligent automation". It's the difference between using the agent as an accelerator of your own judgment versus delegating the judgment itself.
In the first case, the human has already mentally designed what should happen. The agent executes faster, but the correctness benchmark exists in the supervisor's mind. Errors are recognisable because there's an internal standard against which to verify.
In the second case, the human asks the agent to decide what to do. There's no internal benchmark. Errors can go unnoticed because there's no mental model of what success should look like.
The difference isn't moral. It's structural. Those with domain expertise can use the agent to amplify that competence. Those without risk delegating decisions they can't evaluate to systems they can't correct.
The regime that's emerging
AI agents don't operate in isolation. Project Mariner navigates the web, interacts with sites, fills forms, executes actions. When millions of agents operate simultaneously in the same digital environment, we're no longer observing human interaction mediated by tools. We're observing interaction between autonomous systems with humans supervising after the fact.
It's a completely different coordination regime. In algorithmic trading, we discovered this the hard way: algorithms had developed patterns of mutual interaction that no single operator had designed or predicted. The emergent system was greater than the sum of its parts, and its properties were opaque to all participants.
With AI agents operating across more varied and interconnected domains, the potential for undesigned emergent dynamics is amplified. This isn't catastrophism. It's observation of how complex systems behave when autonomous components multiply.
Signals to watch
If this analysis is correct, over the next 18-24 months we should observe some specific patterns.
First: players building agentic infrastructure for internal, invisible processes will show measurable competitive advantage. Not because AI is magical, but because they'll have automated operational bottlenecks while maintaining human judgment at critical decision nodes.
Second: the first cases of "cascade failure" will emerge, where chained agent actions produce results no single user had explicitly authorised. Not necessarily catastrophic, but sufficient to reveal gaps in the supervision regime.
Third: the conversation about responsibility will shift from "who supervises" to "what does supervision mean". It's easy to say humans must control. It's far less clear what this means when the agent operates at speeds and complexity that exceed real-time verification capacity.
If instead we observe widespread adoption without significant incidents and without competitive bifurcation, this analysis is falsified. Agents would then be more robust and less problematic than historical precedents suggest.
The question that remains open
The "agentic era" framing assumes that the transition from tool to agent is a natural and desirable evolution. This assumption deserves scrutiny.
Some problems benefit from end-to-end automation. Others require human judgment distributed at specific points in the process. The temptation to automate everything automatable isn't new. Organisational history is full of systems that removed human discretion in the name of efficiency, only to discover that discretion served to handle exceptions the system couldn't predict.
More capable AI agents might handle more exceptions. Or they might simply fail in more elaborate and harder-to-diagnose ways.
Anyone adopting these systems would do well to ask not only "what can the agent do" but "what happens when the agent fails in a way I hadn't anticipated". The answer to this question determines how much error margin the system can tolerate. And how much error margin the supervisor can afford.