The problem

Every week there’s another talk, another YouTube video, another arXiv paper about AI agents, and they all converge on the same worry from different angles: compute is basically free now, agents can run overnight, fan out across worktrees, loop until a test passes — but the human directing them hasn’t gotten any faster. I kept noticing the same pattern across completely unrelated sources: burnout from running too many parallel agent sessions, teams shipping more PRs with less review, engineers who can’t explain what their own codebase does anymore. Nobody was connecting these into one thesis. At a company this is not a personal productivity complaint — it’s the same failure mode that shows up as a two-tier engineering team (a few people generating 10 PRs a day while everyone else drowns reviewing their output), or as a 242% jump in incidents per PR industry-wide as review became the bottleneck nobody redesigned for.

The approach

This is a live research project inside CogitOS, my personal knowledge-capture system, not a piece of software with its own screen. The mechanics: whenever I watch a talk or read an article that’s relevant, I capture it as a dated idea entry with a full paraphrase and a source link, tag it to this project, and it lands in a per-project folder (manifest.yaml for the project’s own thesis statement, ideas.md as the running log, refs.md and per-source bundles for the original material). Right now the human-attention project has pulled together 24 sources — conference talks (AI Engineer, WorkOS, Notion, Amazon AGI Lab, KRAFTON), YouTube commentary, an arXiv literature review of 46 papers on GenAI in software architecture — into 20-plus distilled ideas, each one tied back to exactly who said it and where.

What it actually does for me: instead of half-remembering “someone said something about review bottlenecks,” I can pull up this project and get the whole argument assembled — Karpathy’s “AI psychosis” framing next to Watershed’s structured-report answer to it next to the Faros AI survey data on review collapse, all cross-referenced. The manifest’s description field is itself a synthesized thesis, not just a list: humans supply judgment and the quality gate, agents supply scale, and the job is routing — deciding which decisions need a human and protecting attention for exactly those, not scaling delegation linearly just because agents are available. The project also has a next_steps slot, empty for now, which is where the research turns into something I actually build or write next — the courses and content pipeline elsewhere in CogitOS reads directly out of folders like this one.

What I learned

The clearest, most repeated finding across sources was that the bottleneck moves, it doesn’t disappear. At low agent capability you read every line of generated code; at today’s capability you route changes to the right verification method instead of reading all of them; at the next tier you stop checking if the work was done right and start checking if it’s the right work at all. Several unrelated speakers landed on structurally the same fix independently — organize review and delegation as a hierarchy (CEO/VP/manager/worker, or planner/worker/judge) rather than a flat pool of agents, because a human coordinating more than three or four peers becomes the scheduler, memory, and reviewer simultaneously and stops scaling. The most useful negative finding: “cognitive surrender” — accepting an agent’s answer as your own without forming independent judgment first — is empirically worse than people assume; a Wharton study found 73% of people kept the wrong AI answer and felt more confident doing it.

Where this could go

The obvious next step for this project is turning the accumulated thesis into something published — a course module or blog series built directly from the ideas log, which is exactly the workflow CogitOS’s course-production pipeline is designed to consume. Beyond my own use, the pattern generalizes cleanly to a team or org: the same attention-routing problem that shows up in one person’s overloaded agent sessions shows up at company scale as engineering-org design — who reviews what, which changes get full scrutiny versus a shadow-mode trace, how you avoid the two-tier team where early adopters ship 10x and everyone else drowns reviewing it. A version of this research discipline — a standing, sourced, continuously-updated thesis file instead of a one-off slide deck — is something I’d build for a client team directly: not a research summary that goes stale, but a living document tied to primary sources that gets revisited every time new evidence shows up.