cd /news/artificial-intelligence/study-an-ask-or-act-tool-availabilit… · home topics artificial-intelligence article
[ARTICLE · art-72147] src=lightningjar.com ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

Study AN: Ask or act: tool availability dissolves the visibility-clause tax

A study by AN found that providing a get_template_view tool eliminated the visibility-clause tax in outline views, raising anaphora success to 45–48 out of 48 with near-zero asks across four arms, three models, and 1,728 session steps. The 'prefer fetching over asking' clause added no benefit where tools existed, and proved unsafe below the frontier: Sonnet silently guessed half of unanswerable cells, Gemini burned 689 tool calls, while Opus asked a perfect 48 out of 48. The clause remains unshipped, with the fence being capability-shaped: give tool-bearing surfaces the ask rule freely, and never pair the rule with skeleton views on a surface lacking tools.

read1 min views1 publishedJul 22, 2026
Study AN: Ask or act: tool availability dissolves the visibility-clause tax
Image: Lightningjar (auto-discovered)

Study AN · Views & grounding

Ask or act

Study AG left one exposure open: under outline views, the shipped ask rule converted fully specified edits into NEED-INFO about 70% of the time, because the target node was "not visible"; even with the last-edit echo supplying id, key, and both values. AN re-ran that construction with a get_template_view tool in hand, four arms, three models, 1,728 session steps.

Availability alone settled it: anaphora success rose to 45–48/48 with essentially zero asks on every tier, models fetching at most one or two precise views or patching directly from the echo.

The registered "prefer fetching over asking" clause added nothing where tools existed; and the protection arm showed it is unsafe below the frontier: told to prefer tools, sonnet silently guessed half of its genuinely unanswerable cells and gemini burned 689 tool calls hunting the conversation-borne referent in the tree, while opus asked a perfect 48/48.

The clause stays unshipped; the fence is capability-shaped: give tool-bearing surfaces the ask rule freely, and never pair the rule with skeleton views on a surface that lacks the tools.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @an 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/study-an-ask-or-act-…] indexed:0 read:1min 2026-07-22 ·