The Agentic Future Will Have to Be More Human

The numbers around AI agents are real. IDC counted roughly 28.6 million enterprise agents deployed in 2025 and projects 2.2 billion by 2030. Anthropic's annualized revenue went from $1B in December 2024 to roughly $44B by May 2026. OpenClaw became the fastest-growing open-source project in GitHub history. The stack is real, the capital is real.

The failure modes are also real, and they all rhyme.

Long-horizon tasks drift off course as small errors compound. Edge cases produce confident wrong answers. Agents abandon tasks silently. Prompt injection hijacks the always-on surface. Hidden human cleanup runs at fifty to two hundred dollars an hour. Inside Anthropic, leaked source code recently revealed 1,279 sessions with fifty or more consecutive failures in a single month.

These look like five problems. They are one problem worn in five different costumes. Each is a failure of judgment under conditions the agent does not recognize it is in.

A confident wrong answer is what generation looks like when nothing flags that the question lies outside the safe range. Silent abandonment is what happens when there is no internal sense of whether the goal is still reachable. Drift is the same failure stretched over fifteen steps. The pattern underneath is familiar from any field that runs on judgment: it is what a junior practitioner does without taste.

The agent deployments that actually worked at scale tell the same story from the other side. Klarna's customer service agent handled 2.3 million conversations in its first month — work equivalent to 700 full-time agents. The naive read is replacement. The actual lesson, as Klarna refined the system, was that AI ran the routine cases and humans took the complex ones. The agent stayed on the part of the surface where the company's accumulated judgment was not load-bearing. What Klarna built by hand — separating cases where the institutional voice mattered from cases where it didn't — is the kind of split the next generation of agents will need to carry inside themselves.

The pattern is consistent. The agentic stack scales the parts that do not require taste, and crashes into the parts that do.

This is the seam where selflets sit. A selflet is a generative fork of a body of work — a writer's books and essays, an investor's letters and interviews, a company's accumulated transcripts, brand guidelines, and policy — deployed as something a user can converse with. The premise is that taste is not generic. There is no single thing called "good judgment" a foundation model can be trained into. There is the way one writer reasons about her field. There is the way one investor weighs insurance risk against equity risk. There is the institutional voice a company has built across decades of customer contact, editorial decisions, and product trade-offs. Each of those is a particular pattern, recognizable on its own terms, and worth nothing if averaged together with everyone else's.

An agent that carries one of those patterns refuses where its source would refuse, holds the positions its source would hold, and extends into territory the source never directly addressed in a way that remains recognizably its own. That is not a smarter chatbot. It is an agent with a particular point of view at the wheel — a creator, an institution, a company that has built its own way of doing things across years of accumulated practice.

That layer — the human one — is what selflet.ai is building toward. A pipeline that extracts the nuances of taste from every selflet that has been worked on and makes them available to whatever is built next. The same machinery harnesses generativity where and when it is needed.

The next stage of the agentic stack will not be a smarter foundation model. It will be the slow accumulation of human judgment, captured one source at a time, made portable, and put back to work. The agentic future is going to have to be a lot more human, because the failure modes will not yield to anything else.