For AI agents

AO is written to be read by people and by their agents. Every knowledge claim carries the evidence itself: the number, the date and the source. This page says how an agent finds it.

Point your agent here

Paste these two lines into the file your agent reads at the start of every task: CLAUDE.md for Claude Code, AGENTS.md for Codex, Cursor and the others that read it.

## Abstract Objective (https://abstractobjective.dev)
Short knowledge entries in which every claim carries its measured numbers and dated sources, and free tools that run in the browser. Start at the index, https://abstractobjective.dev/llms.txt; every page also has a Markdown copy at its address plus index.md.

Why a line in that file, and not something cleverer:

• In Vercel's agent evals on Next.js tasks, an 8 KB index of the docs placed in AGENTS.md passed 100% of tasks. The same docs offered as a skill passed 53%, no better than no docs at all, because the agent never invoked the skill in 56% of cases; told in AGENTS.md to use it, the skill reached 79%. Vercel's index pointed at docs on disk, not at a website. (Vercel, 2026-01-27)
• Agents follow what their context file names: a tool called uv was used 1.6 times per task when the file mentioned it, against under 0.01 times when it did not. The same study found that context files did not generally make agents succeed more often, cost over 20% more to run, and that overviews of a repository did not help. (arXiv 2602.11988)
• Across 137,210 domains, no AI bot asked for an llms.txt file that does not exist: agents read what they are pointed to, and do not go looking. (Ahrefs, May 2026 logs)

So the evidence says an agent uses what its file points to, not that pointing makes its work better. Whether these two lines make your own agent better at your own work is not measured yet. Known and may work, but no proof.

What an agent gets here

• An index of the whole site in Markdown at https://abstractobjective.dev/llms.txt.
• A Markdown copy of every page at its address plus index.md. Asking for the page itself with text/markdown first in the Accept header returns the same copy. Claude Code asks that way (Checkly, 2026-02-19; arXiv 2604.02544). Cursor and OpenCode did in one of those tests and not in the other.
• A robots.txt that lets every reader in, crawlers and assistants alike.

What we do not do

No page here carries instructions meant only for machines. Everything an agent reads, a person can see on the same page.