Algorithm & Blues
AI research translated into decisions executives can actually make. One clear argument per issue, published weekly.
Get the weekly issue
Published every Sunday since May 2025. No hype, no filler, one clear argument per issue.
The One-Person Team
Solo-authored research papers began rising again around ChatGPT's release, especially in fields where work happens on a computer. Akira Matsui's analysis suggests AI may let researchers carry familiar, narrower work without coauthors. Companies may see the same effect as one person moves faster across work that once required several roles. The saved coordination time is real. So are the conversations, context, and junior assignments that disappear when the handoff does.
Hotel expense reports are a natural job for an AI agent, and Microsoft researchers gave one to a GPT-5 agent inside Dynamics 365. It got better when they took most of its history away: completion rose from 71 to 91.6 percent once older work was replaced with a short summary, while tokens and runtime fell about 60 percent. The failure it kept hitting was stale state, acting on an outdated version of the form when several were in context at once. A larger context window gives an agent more room, but it doesn’t decide what to forget.
Imagine testing a product change on eight billion people without speaking to any of them. A new preprint, MatrAIx, gets close, simulating a public from 8.3 billion persona records that answer surveys, use apps, and react to price changes. But swap the model playing the customers and the market conclusions can flip, and the reliability gains lean on an assumption of independence these agents rarely have. The synthetic public is cheap, fast, and endlessly cooperative, which is exactly what makes it easy to mistake for the real one.
Multi-agent systems promise a built-in second opinion: one agent plans, another executes, a third checks. A new paper by Ruiwu Niu, Xincheng Shu, and Ying Zhao finds the catch. The connectivity that lets agent review improve the majority vote is the same connectivity that lets a false claim spread, and sometimes no network design satisfies both. Worse, five approvals can trace back to a single source, which turns AI review into a lineage and segregation-of-duties problem rather than a genuine independent control.