How Proofable uses Context.dev to prove what every agent read on the web
Agents browse the web, buy, post and send email for the people who run them, and every action can leave a receipt any app or agent can check later.
That work starts with what an agent reads. A bad page becomes a bad decision, and everything after it inherits the mistake.
What Proofable runs on Context.dev
Web search and page reads run on Context.dev for scraping, search and crawling through one API, with JavaScript rendering and proxies handled upstream, so we run no scraper fleet of our own.
We moved there over a weekend in July, after raw HTML was flooding agent context with noise.
What a proven read changes
An agent pins what it read, and the result it produced, into one signed record. The next agent checks that record before it acts, so a page that changed between reads and a result altered in transit both become visible. The forwarded text is never taken on trust.
What we measured
We read one URL twice, 221 seconds apart. The first read returned 96,713 characters. The second returned 1,129. Both reads succeeded, and neither agent could tell the difference.
When we altered a result passed between two agents, the agent that trusted the text used the wrong value 3 of 3 times. The agent that checked the record caught it 3 of 3 times.
On six public questions about standards and regulation, reading signed records halved the cost per correct answer and cut worst-case token spend to 16,301 from 50,554, with a checkable source cited on 6 of 6 answers against 5 of 6 on the open web.
Both arms answered 6 of 6 correctly, so this measures cost, worst-case spend and provenance, not accuracy.
Where the two products meet
News search reaches more than 200 million articles, Highlights returns the passages that matter instead of the whole page, and Monitors already sends a signed webhook when a watched page changes. A pinned read is the same idea one step earlier in the run.
The impact
-
Agents stop building on data they cannot prove.
-
One provider for search and page reads.
-
Every web read carries a signed receipt the next agent can check.
-
A changed page or an altered result gets caught instead of inherited.
Check it yourself
Every run live, with the proof: docs.proofable.me/evidence
Building agents that need to trust what they read?
Start free at proofable.me/