Tool Crowding Benchmark

When an agent has dozens of MCP tools loaded, does it start picking the wrong one? Tool Crowding is a harness built to measure that cleanly, with the design and its hypotheses pre-registered before any data.

Every trial is cache-cold: a per-trial nonce defeats prompt caching, a runtime assertion halts the run on a cache hit, and server SHAs plus tool schemas hash into every run ID, so a result can be reproduced byte for byte.

Status: harness built, 19 exploratory trials run, the 144-trial pilot not run yet. The exploratory trials found no mis-routing with 6 dissimilar tools, which points at task ambiguity and agent persona rather than raw tool count.