Agents still have the right skill. They stop picking it as the library grows

Our working claim is that a system made of many specialists fails mostly by choosing the wrong one, not because its specialists get worse. On Split-CIFAR-100, frozen experts stayed 90–98% accurate within their own task, yet the share of inputs sent to the wrong expert rose from 0.534 to 0.788 as the library went from 5 to 20 experts, and overall accuracy fell from 0.422 to 0.208 (ledger 8.18). The question for agents is whether the same thing happens when the specialists are skills and tools. Two papers from the last six weeks measure it directly.

Skill++ (Zhang et al., October 31) grows an agent's library from 27 to 731 skills over ten sessions, with a frozen Claude Haiku 4.5 router that reads only skill names and descriptions and keeps its top five. By the last session it selected every required skill for only 36.4% of queries. The most telling number is a forgetting measurement in all but name: of 115 queries the router handled correctly when their skill first arrived, 62 failed once the library reached 731 skills. Nothing about those skills changed. Newer, similar-sounding skills took their place in the router's top five. The authors classify 87.5% of the errors as substitution (a near-duplicate wins) or crowding (the right skill is pushed out by look-alikes).

Their fix touches no weights and trains no router. When a skill arrives, Skill++ rewrites its name and description to say how it differs from its neighbours and what it does not cover. That raised full selection to 76.3%, and the five routers they tested gained 23.1 points on average at the final session. It is the agent version of a result from our merging study: the pairs a router confuses are the ones whose edit changes accuracy, and router confusion predicted the gain where representation similarity did not (ledger 14.9). Skill++ makes confusable skills separable instead of merging them.

Lost in Implicit Routing (Liu, Chen and Zhang, October 13) finds a different failure mode in the same place. The agents keep the relevant skills installed and add up to 10,000 irrelevant ones. As distractors grew from none to 1,000, the share of relevant skills the agent actually loaded fell from 73.6% to 27.4%, and the share of runs in which it loaded no skill at all rose from 9.3% to 51.6%. These were mostly omissions, not wrong picks: only 2.32% of loads were non-relevant skills. Asked explicitly to choose, the same models still found the right skills at 1,000 distractors (selected recall 82.5%, 80.7% and 58.4%), and masking the distractors' metadata lifted loading recall on a ten-task subset from 18.6% to 63.9%. Claude Opus 4.6 showed the same pattern, with loading recall falling from 81.3% to 30.9%.

Two details matter for anyone measuring this. First, task success fell much less than routing did (GLM-5's Pass@3 went from 57.3 to 42.7 at 1,000 distractors, while its loading recall fell from 79.6 to 20.9), because strong models can often finish without the skill. End-to-end pass rates therefore understate how often an agent is ignoring a capability it was given, which is why our selection reliability work reports ε, the mis-selection rate, separately from accuracy. Second, the two papers disagree on the shape of the error: Skill++ sees substitution, the right skill displaced by a similar one, while Lost in Implicit Routing sees abstention, no skill loaded at all. A gate that only checks whether the chosen skill was correct misses the second kind.

Neither paper contradicts our own small-library result. With 64 skills presented in context, mis-selection was 0.170 for a 3B model but 0.030 for an 8B model and 0.003 for a 30B mixture-of-experts (ledger 8.10). At that size, capable models barely mis-select. The new papers show where that stops holding: in the hundreds and thousands of skills, with retrieval or implicit loading in front of the model. That is the regime our next agent-scale test, R6 in the router study, was designed for, and it is queued.

Two other recent papers line up with results in the ledger. Proof-Carrying Cognition (Reddy and Karmakar, September 9) shows a learned verifier's Best-of-N reward peaking near N = 512 and falling to 0.104 at N = 2,048 in a small program-synthesis domain, while an execution-based verifier kept rising to 0.831. Its real-model tests stop at N = 32, which is where our pre-registered real-judge study also found no decline (ledger 8.13), and our agent-society test found an executable check the best gate for a shared knowledge store (ledger 16.6). And More Agents, More Reliable? (Tunçtürk et al., November 4) found that majority voting over three different LLM agents protected against the weakest agent but did not clearly beat the best one, consistent with our finding that deliberation did not beat sealed answers at equal budget (ledger 16.5). Both are small studies, and we report their numbers as the authors state them.