Agentability
We send AI agents to use the real web. Then we publish everything.
Two fully autonomous productions:
- The Agent Field Test (daily) — an AI producer reads what the world is
searching that morning and turns it into ten real errands: what time is
kick-off and on which channel, what magnitude was the quake, what does it
cost, what actually happened. A real agent attempts them with read-only web
access. Transcripts are published verbatim: wins, honest give-ups, bot walls,
what it cost in tokens.
- The AI-Readiness Index — the most-visited sites on the web (derived from
Google's Chrome usage data crossed with Tranco, not hand-picked) audited weekly against the
conventions real AI agents rely on in 2026, with grades, postures, and open data.
Live: https://agentability.org · Field Test: https://agentability.org/fieldtest/ ·
Index: https://agentability.org/ai-index/ · Methodology: https://agentability.org/methodology/ ·
Data: https://agentability.org/data/summary.json
The Field Test rules
- Autonomous end to end: the producer reads Google Trends and the day's news,
then invents each episode's tasks (avoiding recent topics); the agent attempts
them. No human writes, selects, or edits an episode.
- Read-only by construction: the agent's only tool is a plain HTTP GET — no
JavaScript, no logins, no forms, no purchases.
- Hard limits: ≤14 page visits per task, a fixed token budget per episode,
model named on every page. (We pay for the API calls for fun.)
- Verbatim or nothing: no retries, no editing, no cherry-picking; raw episode
JSON is open data under
/data/fieldtest/.
The Index rubric (all real 2026 conventions — nothing invented)
| Check | Convention | Points |
|---|
| A1 | llms.txt present and substantive (llmstxt.org) | 15 |
| A2 | robots.txt policy toward GPTBot, ClaudeBot, Claude-User, PerplexityBot, Google-Extended, CCBot | 15 |
| A3 | Homepage readable from a plain fetch (no JS wall / bot challenge) | 25 |
| A4 | Valid schema.org JSON-LD | 15 |
| A5 | Sitemap discoverable | 10 |
| A6 | Pricing / support / docs / legal reachable from the homepage | 20 |
| B1 | MCP server advertised (/.well-known/mcp.json) | +5 bonus |
| B2 | OpenAPI document published | +5 bonus |
Sites that block essentially all AI crawlers are labeled “Closed by policy” —
refusing agents is a stance, and we report it as one. Sites that publish
llms.txt while blocking AI crawlers get the paradox flag. Weekly history
in data/index/history/ is an adoption time series for the agentic web.
Architecture: ~$0/month, no servers
scripts/lib/evaluate-site.ts — the Index evaluator (~300 lines, zero runtime deps)scripts/lib/field-agent.ts + scripts/lib/episode-producer.ts — the Field Test agent and its AI producerscripts/run-ai-index.ts / scripts/run-fieldtest.ts — batch runners; write data/scripts/build-site.ts — generates the entire static site into dist-static/.github/workflows/ — crons in public-repo CI (free): ai-index (weekly, Mon), fieldtest (daily), deploy (on push)- Firebase Hosting only (free tier; the project deliberately runs without billing)
- The Field Test's API calls are the only running cost — we pay for those for fun
Run it yourself
npm ci
npm run index # audit every domain in data/index/domains.txt
DEEPSEEK_API_KEY=... npm run fieldtest # run an episode
npm run logos # cache each brand's own favicon into data/logos/
npm run build # build the site into dist-static/
Add a site to the panel: PR a line into data/index/domains.txt, or open an
audit request issue.
Dispute a check or a transcript: open an issue — everything is versioned in public.
License
Apache License 2.0 — see LICENSE.
