Browser Tool Action Inventory
Give browsing agents a small closed command set — search, click-by-ID, find, quote, scroll, back, end — instead of free-form browser scripting.
WebGPT’s ACI exposes a fixed command vocabulary: Search, Clicked on link
<link ID>, Find in page, Quote, scroll up/down by N, Top, Back, and
End: Answer / End: Nonsense|Controversial. Observations use
lm-oriented web observations with stable
link IDs so click targets are parseable. Prefer a modern search API so the LM
practices using a search engine, not reimplementing retrieval indices
(retrieval-augmented generation remains
the alternative when you own the corpus).
Keep the inventory small under ACI design principles and tool inventory. Treat non-matching generations as invalid actions that still spend budget. Episode shape often splits into browse-then-answer phases.
Sources
- WebGPT (Nakano et al.), §2 Table 1