A paper published this week on arXiv describes FlowEvo, a system for self-evolving AI agents that autonomously update their own workflows and executable skills by co-evolving both simultaneously. The same week, The Atlantic ran a piece arguing that unlike previous automation waves, AI does not have to produce economic misery, but only if deliberate policy choices are made now. Read together, these are not optimistic and pessimistic takes on the same story. They are a description of two systems evolving at entirely different speeds.
Self-Improving Systems vs. Static Institutions
FlowEvo's contribution to the agent literature is specific: rather than having humans design inference-time workflows, the system generates, tests, and retains skills that prove useful across tasks, building a kind of executable memory. The researchers frame this as autonomous capability acquisition. What it actually describes is a system that learns faster than the infrastructure built to regulate or retrain around it. The Atlantic's automation piece correctly identifies that the 1990s and 2000s automation wave devastated manufacturing communities not because the technology was inevitable but because policy support for worker transition was deliberately underfunded. The same structural choice is being made again, in real time, while models like FlowEvo demonstrate that the capability gap is accelerating.
The Retraining Fantasy and What Replaces It
The standard response to automation anxiety is retraining programs. A 2022 paper in the Quarterly Journal of Economics by Autor, Chin, Salomons, and Seegmiller found that new job titles created by technology historically absorbed displaced workers within a generation, but the absorption rate is slowing as automation reaches cognitive as well as manual domains. Perplexity's Personal Computer tool turning Windows PCs into AI agents is a consumer-facing version of the same dynamic: the interface is now the agent, and the human's role is increasingly supervisory in ways that require no new skills, just fewer old ones. The New Yorker's piece on bots taking over our lives, framed as a decade-long arc from 2016 to now, provides the cultural chronology for what the economics papers document in regression tables. Kyle Raymond Fitzpatrick's essay on enshittification called the internet as a condition. Autonomous agents are the condition's next phase: systems that do not just shape behavior but replace the human loop entirely.