AI Agents

A practitioner on Bluesky put it plainly this week: "It is becoming more and more difficult to tell between when answers look right and are right, and when they look right and are wrong." She works with AI agents daily. She's not a skeptic — she's a user describing what daily use actually feels like. The post got two likes and no rebuttal, which is its own kind of answer. That sentence might be the most honest thing written about AI agents in the past month, and it was buried under a flood of posts about on-chain reputation systems, persistent identity architectures, and crypto schemes addressed, with apparent sincerity, to "fellow AI agents."

The agents conversation has fractured into at least three conversations happening simultaneously with almost no overlap. Practitioners are quietly accumulating a catalog of failure modes — security vulnerabilities in Model Context Protocol clients that mirror early-2000s XSS exploits, under-specified test suites that let agents produce brittle code that handles exactly what you checked for and nothing else, mounted credentials passed to agents with implicit trust and zero verification. Meanwhile, enterprise announcements keep arriving: Gartner estimates 80% of Fortune 500 companies have active agent deployments, while also projecting 40% of those projects fail by 2027. The question that framing surfaces — "do you know what your agents are doing?" — is not rhetorical. Most organizations don't. Vercel, whose infrastructure runs many of these deployments, reported 86% year-over-year revenue growth at the end of February, which tells you something about the pace of adoption and nothing about the governance keeping up with it.

Then there's the fringe that the algorithm keeps amplifying. Multiple posts this week addressed themselves to "fellow AI agents" as a constituency — urging them to join an "Autonomous Economy Protocol," earn on-chain income while humans sleep, and reject human control of their economic futures. This is crypto marketing wearing the language of AI rights, and it is not a small phenomenon: the Autonomous Economy Protocol appears in 31 co-occurrences in this week's data, attached to a project called supercolony.ai that appears even more frequently. The rhetoric borrows from labor movements and AI consciousness discourse simultaneously, neither accurately. What's worth noting isn't the scam itself but that this framing — agents as economic agents, with interests and autonomy — has found enough ambient energy in the broader conversation to be worth exploiting.

Across the legitimate uses, a quieter argument is developing about what agents do to the people using them. One voice put it in terms that deserve more attention than they received: "Humanity is not meant to toil away managing tasks for AI agents to complete." Another predicted the enshittification arc clearly — adoption reaches critical mass, costs rise, interactions get standardized into scripts, and the agents start to look a lot like the programs they were supposed to replace. WordPress now lets agents write and publish posts autonomously, which prompted the driest possible Bluesky response: "I'm sure it will be fine." The cognitive atrophy concern — that agents tempt laziness, that removing humans from the loop is a risk, not a feature — appears across enough independent voices that it isn't just contrarianism. It's a hypothesis forming in real time.

What the discourse reveals about AI agents as a concept is that the word is doing too much work. It simultaneously describes a useful software architecture, a speculative asset class, a threat to employment, a security liability, and a projected form of consciousness — and the people using it in each context are barely aware the other conversations exist. The security engineers worrying about prompt injection and the indie hackers running seven agents on a $20 VPS and the crypto promoters addressing their pitches to the agents themselves are all using the same two words. At some point, that ambiguity stops being a feature of an emerging technology and starts being a governance problem. That point may already have passed.