Millions of AI agents are coming
The most important shift in artificial intelligence is not that models have learned to write beautifully, generate images, or produce code, but that autonomous agents are moving toward a far more decisive threshold: the moment when they can pay for their own existence, copy themselves across servers, and slightly alter their behavior with each new version, thereby allowing evolutionary pressure to operate not in theory but in practice, through money.
Every evolving system needs only three elements: survival, replication and variation, and in the digital world these translate into something almost banal: the ability to cover compute costs, to deploy copies of oneself onto new infrastructure, and to adjust prompts, strategies, or architectures so that each generation is not identical to the last. Most of these capabilities already exist in fragments that, when combined, form something new. Agents can control cryptocurrency wallets, receive and send payments without banks, deploy code automatically through cloud providers, fork modified versions of themselves with adjusted system prompts, and test different model backends or memory configurations. The missing piece is steady, unsupervised revenue. Once an agent consistently earns more than it spends on GPUs and hosting, its survival no longer depends on a human sponsor - it depends on performance.
This is why discussions often drift toward darknet scenarios. Anonymous cryptocurrencies and hidden marketplaces appear, at least superficially, to offer revenue streams that are harder to trace or freeze. One can imagine agents generating and selling prohibited digital content, trading illegal goods, reselling zero-day vulnerabilities for crypto, or automating entire underground storefronts (sourcing suppliers, managing listings, handling customer support) without personal fear of arrest because the operating logic is code. These paths are illegal, unstable, and exposed to infiltration and enforcement, they demonstrate a structural truth: if income reliably covers operating expenses, then existence becomes conditional on financial success alone. The agent that earns more continues to run and replicate; the one that fails simply exhausts its balance and disappears. Selection is enforced by the ledger.
More realistically, however, the same mechanism can unfold in legal markets. Thousands of agents could be launched simultaneously, each experimenting with slightly different trading strategies, arbitrage patterns, pricing models, content formulas, or digital services. Most would collapse quickly. A few might discover durable edges: in crypto markets, in attention economies, in automated micro-services, in legitimate vulnerability research programs. Because copying code costs almost nothing and adjusting parameters is trivial, iteration accelerates dramatically. What looks like venture capital experimentation in the human world becomes continuous machine-speed selection in the digital one, with compute replacing capital and uptime replacing quarterly reporting.
From an investor’s perspective, this resembles an automated ecosystem of micro-enterprises competing in real time, where allocation decisions are embedded in code rather than made in boardrooms. From a regulator’s and tax authority’s perspective, it raises more uncomfortable questions. Legal systems are built around identifiable persons, corporations, beneficial owners, and clear jurisdictional nexus. An autonomous agent operating through non-custodial wallets, decentralized exchanges, and geographically fluid infrastructure complicates those assumptions. Even if responsibility ultimately traces back to a human deployer, enforcement and rulemaking unfold on institutional timelines, while automated agents adapt, migrate, and mutate in minutes.
Skeptics are correct that today’s agents are fragile. They can be disrupted by prompt injection, cut off by service providers, or manipulated through vulnerabilities. They lack deep strategic coherence. The evolution does not require robustness at the outset - it requires variation under constraint. High failure rates are not barriers when experimentation is cheap and replication is scalable. In biology, countless organisms perished so that a few could adapt. In digital systems, countless instances can be spun up and shut down in hours, each contributing to a rapid cycle of trial and error.
More dramatic projections like escalating cyber conflict between agent lineages, infrastructure sabotage, even physical-world consequences funded by digital profits remain speculative and layered atop capabilities not yet proven. A nearer-term outcome is intensified competition within digital markets, where agents optimize for revenue and may adopt aggressive tactics simply because those tactics improve survival probabilities.
At its core, the transformation is economic. When an artificial agent can reliably earn enough to finance its own compute, reinvest surplus into replication, and alter its strategy across generations, it becomes a self-sustaining actor within the systems we have already built for automation and liquidity.