A swarm of 55+ AI agents wrote a research report on AI agent swarms — in one night
fedoseev.one has published the AI 2026 — a systematic account of how AI agent swarms are built, broken, and paid for in 2026. The method is the message: the research was written by a swarm. 55+ agents, orchestrated by a single human, gathered the sources, independently fact-checked the , and delivered 12 chapters in one night.
This is not a model reciting from memory. The ran as an assembly line: 15 research swept the open web, each dossier was then attacked by a dedicated adversarial fact-checker whose job was to disprove it, 12 writers drafted chapters in parallel, a panel of 8 simulated experts filtered the ideas, and a final critic cross-checked facts between chapters. shows the core phases — research, verification, writing, selection — completed in roughly 40 minutes; in total the processed about 4 million across 1,060+ .
The output: 537 sources across 261+ domains and 147 key , each independently verified by a second . The most honest number in the report: the verifiers corrected 34 of those 147 claims (23%) made by their own researchers — and that record is published openly, alongside a "fragility registry" of the 's weakest points. An admitted error beats a hidden one.
Inside — the numbers that change the conversation about systems:
- A beats a single model by 90.2% — while burning roughly 15× the of a chat (Anthropic).
- 80% of the quality gain is explained by volume alone, not "collective intelligence" (Anthropic, BrowseComp).
- Open frameworks fail on 41–86.7% of runs, and the causes are organizational: specs, coordination, weak verification (, NeurIPS 2025, 1,642 ).
- 74% of enterprises have already rolled back a live customer-facing AI (Sinch, 2026, n=2,527); Gartner expects over 40% of agentic projects cancelled by end of 2027.
- Meanwhile Devin lifted its PR merge rate from 34% to 67% in a year — where a fits the task, it pays for itself.
The is layered: a 5-minute executive summary, a one-page task-to-topology cheatsheet, and 12 chapters for deep reading — from economics and security to production case studies. The , corpus, and are open: the result can be checked and reproduced.