Latest

Anthropic’s Warning: AI Nears Self-Improvement – Are We Ready?

Anthropic’s AI warning: autonomous agents could speed up crypto adoption

Anthropic says AI may be close to improving itself with less human help. Autonomous agents could build parts of their own systems, test them, then revise the result. Sounds theatrical. It is not hard to see the market version, though: agents that make payments, move assets, and operate software without much supervision would put more load on crypto networks. My take: ETH and SOL may start trading less like abstract AI-adjacent bets and more like infrastructure exposed to agent-driven transaction demand.

Anthropic's Warning: AI Nears Self-Improvement – Are We Ready?

The US AI company made the case in a blog post published Thursday. Jack Clark, Anthropic’s co-founder, and Marina Favaro wrote that AI agents can already run code and hand tasks to other agents. They said the systems are “on the cusp of taking over completely.” For most of AI’s history, humans pushed models forward by doing the hard parts themselves. Now Anthropic says it is handing more AI development work to AI systems. Why does this matter? Because that shortens the feedback loop. Add enough compute and the uncomfortable endpoint is a system that can design and build its own successor.

This is not a distant thought experiment. OpenAI is already researching “recursive self-improvement” as part of its safety work, and it is hiring in that area. The pace is hard to ignore: AI model improvement has been roughly doubling every four months, not every seven. Anthropic says Claude now writes about 80% of the code merged into its own codebase. I’ll be honest: that number changes the tone of the whole discussion. Favaro and Clark still add a warning: “recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for.” Most guides frame this as a productivity story. That’s only half right. Humans may stop writing most code and spend more time reviewing it, which sounds efficient until the review queue becomes the bottleneck. It breaks there.

For crypto, this is an adoption signal worth watching. Imagine billions of AI agents managing digital assets, placing trades, then settling payments for users. Circle CEO Jeremy Allaire predicted in January that this could happen within five years. Some of it is already here. Keyrock said last month that AI-agent payments moved “from concept to reality” over the past 12 months, with $73 million settled across 176 million transactions. The $73 million figure is small by crypto market standards; the 176 million transactions are harder to shrug off. Counter to the usual advice, I would watch transaction count before headline dollar volume here. Automated activity can be tiny per payment and still matter if it repeats constantly. Ethereum and Solana are the obvious names. More block space demand could mean higher fees, more network revenue, plus firmer support for native tokens if the activity lasts.

The story is not just about transaction counts. Anthropic chose not to release its Claude Mythos model because of cybersecurity concerns. The company said the model could “easily create software exploits.” That is the rough side. The same coding ability could also help developers build better decentralized apps and protocols in controlled settings. I would not treat that as guaranteed. Still, it is believable. Leaders at Anthropic, OpenAI, and other AI companies want stronger guardrails, especially for risks like biological weapons. Here is the contradiction: stronger guardrails may be necessary, but they probably will not slow the race much. Companies are competing with each other. Governments are competing too. Without global coordination, a real slowdown looks unlikely, which means more capable agents could arrive sooner than many people expect. Crypto may absorb part of that activity through automated finance.

What this means

AI development is moving toward agents that can act without constant human direction. For crypto, that could become a real source of network activity, not just another tech narrative looking for a chart. Keyrock’s $73 million in AI-agent settlements is still early, but 176 million transactions gives a cleaner picture of scale. Is this overkill for investors to track now? For a market that reprices narratives fast, no. If agent payments grow, blockchains will need to handle small, frequent automated transactions at volume. That favors low-fee networks with high throughput. It could also lift transaction counts and total value locked across smart contract platforms, which would matter for tokens like ETH and SOL if demand for block space rises.

Investors should watch AI apps and protocols built around automated finance. Agent-to-agent payments deserve their own bucket. The useful metrics are concrete: AI-agent transaction growth by chain, new protocols built for agent interaction, plus blockchain integration news from major AI firms. I would keep Circle’s quarterly reports on the desk too, since stablecoin volume could rise if AI agents settle more payments. Regulation is the pressure point that can flip the setup quickly. Serious debate about autonomous agents in financial markets could change risk appetite fast. The next major update from Anthropic or OpenAI, or a major AI conference with hard numbers on agents, could move crypto markets, especially tokens tied to fast, low-latency chains.