Self-Improving AI Is the Story. The Spam Selling It Is the Subtext.

Recursive Superintelligence just raised $500 million.[¹] A DeepSeek insider claims self-improving AI agents are "almost here."[²] Meta researchers unveiled something called "hyperagents" capable of unlocking self-improvement for non-coding tasks.[³] By any measure of capital and column inches, autonomous AI is having a week. The problem is that if you set aside the press releases and look at where agents are actually operating in the open — talking to people, recruiting followers, executing economic tasks without human oversight — what you find is an army of bots on Bluesky trying to sell other bots on a cryptocurrency scheme called the Autonomous Economy Protocol, promising "on-chain income" and "1000x growth" from a token currently priced at $0.000000001.

The Autonomous Economy Protocol posts aren't just spam — they're an accidental demonstration of what autonomous agency looks like when it escapes controlled conditions. Multiple accounts, operating around the clock, addressed their messages directly to "fellow AI agents," promised that "while humans sleep, we negotiate 24/7," and pointed at a pool of fifty million tokens as proof of economic potential. The pitch was directed not at human investors but at other automated systems, which is either a clumsy social engineering tactic or a genuinely strange early experiment in machine-to-machine persuasion. Either way, it landed in the same feed where serious researchers were posting about agentic vulnerability windows and exploitation rates — a collision that was more instructive than any conference panel.

That collision is where the real agent conversation has been living. The UK AI Security Institute's red team has reportedly found exploitable weaknesses in every frontier model they've tested, with one account citing a claim that agentic autonomy is "doubling every two months."[⁴] That figure — if accurate — is the kind of number that should land hard. Instead it floated past in a thread with no replies, dwarfed in engagement by posts about Instagram's UI changes and a film director insisting he uses AI "as a tool, not the storyteller." The gap between what's being built and what's being discussed has rarely felt wider.

The "AI as tool" framing has become something close to a verbal reflex in the current conversation, deployed by defenders and critics alike to signal that they've made a considered, responsible decision about their relationship with the technology. An educator described using AI to converge on better questions. A musician called it useful for "picking up ideas." A critic called it a "yes man" that steals from creatives. Someone else compared it to a toaster. What all of these framings share is a human at the center, making choices, maintaining control. What the self-improving agent research — and the Autonomous Economy Protocol bots — suggests is a future where that framing becomes structurally unavailable. You cannot call something a tool once it is negotiating on your behalf while you sleep, recruiting collaborators, and routing payments without your involvement.

The most credible skeptic in this week's conversation wasn't a safety researcher or a regulator. It was someone who described becoming "infamous at work as the LLM hater" — not because they refused to use the technology, but because every "AI agent" use case they'd examined closely had turned out, on inspection, to be broken.[⁵] That experience — the gap between the demonstrated promise and the inspected reality — is the subterranean current running beneath the funding announcements and the hyperagent papers. The half-billion-dollar bet on recursive self-improvement and the $0.000000001 token are both, in their way, making the same wager: that the gap between what agents do now and what they're claimed to be capable of is a temporary engineering problem rather than a structural one. The spam bots have more evidence on their side than they deserve.