Skip to content
AIOS

Research Wiki

Most of the writing about working with AI is about the machine — better prompts, better context, better tools. This wiki is about the human: what decades of human-factors research already know about attention, vigilance, automation, and supervision, and what that means now that a developer’s job is increasingly to supervise agents rather than write code by hand.

It exists because the same failure modes that human-factors researchers documented in aviation, process control, and cognitive psychology — vigilance decrement, attention residue, the ironies of automation — reappear, almost unchanged, when one person orchestrates several coding agents at once. The canon is old; the application is new.

What it does

Section titled “What it does”

Three jobs, in priority order — the same shape as the Agentic Engineering knowledge base, pointed at the human side:

  1. Surface the highest-signal research so a practitioner can reason about their own attention as deliberately as they reason about a model’s context window.
  2. Store durable summaries of the canonical works, so the insight survives paywalls and link-rot — each with the one reusable technique and what it means for supervising agents.
  3. Feed the Agentic Maturity model — specifically the candidate cognitive-ergonomics lens — and the workspace’s own attention instrumentation.

It is not a bookmark dump. Every entry earns its place by connecting to an AM level, axis, or pattern, and by answering a practical question: how does this change the way I work with agents?

How it’s organized

Section titled “How it’s organized”
Sources One page per canonical work — a durable summary, the reusable technique, a verifiable quote, and what it means for agentic engineering.
People The researchers whose work grounds the human-factors canon.
Cognitive Load The lens: attention residue, vigilance decrement, supervision fatigue — and how the workspace measures them.
Playbooks End-to-end, attention-protecting workflows you can run today.

How an entry earns its place

Section titled “How an entry earns its place”

Adapting the wiki blueprint that governs the pattern library, every sources/ entry carries a small, visible metadata block — author, year, type, the AM axis it informs, tags, and an actionable rating — and every entry answers the same closing question: what this means for agentic engineering. A summary that cannot connect a finding to a concrete change in how you supervise agents does not belong here yet.