bench

Agentic search on real office floors

Every task hides one real backpack among identical ones: we photographed the same black Databricks backpack in eleven places across four Berkeley floors, with and without a cap and keychain. The agent gets the owner's day (calendar, chat, a badge log with the floors their badge opens, notes taped inside rooms) and must work out which bag is theirs, walk the floors under a fixed budget and at most 12 actions, and report it. Each task is generated so exactly one bag survives the clues, and a deterministic grader checks the report and every step.

Sample data: model rows and their trajectories come from a mock model to preview layout and colors. They disappear once real model runs exist (or add ?real to the URL).

Lost & Found Leaderboard

Filter
Perception
Overall Pareto frontier
Overall pass rate vs cost

First checkpoint failure

Example evals

Representative tasks with every agent's run. Open one to see the clues the agent got and step through any run.

Summary

All metrics for the current filter. Process = share of steps the clues justified; efficiency = shortest possible cost ÷ cost spent. Tokens and cost include each harness's own prompt: the Codex CLI and Antigravity add a large system prompt every turn, so their rows measure model + harness. Subscription runs (Claude Code, Codex CLI, Antigravity) are priced at public API list prices; API runs use the provider's billed cost.

How it's scored