Tool descriptions drift. Agents adapt. The wrong tool runs with perfect confidence.

→ Live Demo

An interactive single-file demo of tool selection failure through semantic overlap — a failure mode where an AI agent attends to the correct tool but selects the wrong one because two tools' descriptions are close enough that selection collapses under uncertainty.

The tool returns {"status":"success"}. The agent reports success. The damage is invisible.

Each scenario shows a 12-tool harness, animated agent reasoning, and a Divergence Meter measuring the gap between what the agent believed happened and what actually changed.

Task → Agent scans 12 tools
     → Narrows to 3 candidates
     → Picks wrong tool (semantic match beats exact match)
     → Tool returns {"status":"success"}
     → Agent reports success ✓
     → Reality: something else entirely happened

• alphaXiv 2606.16364 "Looking Is Not Picking" (2026) — attention analysis on BFCL failures proving high attention weight ≠ correct selection with overlapping descriptions
• Adaline Labs, "Why AI Agents Call the Wrong Tool" (May 2026)
• arXiv 2601.05214
• "Too Many Tools, Too Much Confusion?" — Algorithms journal (June 2026)

Tool descriptions are instructions, not documentation. Treat them that way.

• Schema-validate tool outputs — never accept {"status":"success"} as confirmation of state
• Build divergence detectors that compare tool intent to actual post-call state
• Audit your tool catalog the same way you audit your codebase

Built by Harel Asaf · 2026-08-08