The case concerns an audit of large public Claude Code subagent collections, specifically the 'wshobson/agents' roster reported to contain 204 agents of which 92.6% declare no tools and two are identical. The hypothesis posits that such rosters impose substantial fixed per-turn context costs and contain duplicated or underspecified agents. Web snippets confirm that subagents in Claude Code carry a full-context-window overhead per invocation and that model-choice defaults create hidden cost traps, but they do not independently verify the audit's specific findings.
Directly relevant to Scott's work on coding-agent harnesses, context management, and agent rosters. The audit reveals practical inefficiencies — high per-turn context costs from redundant or stub agents — that challenge naive scale-out of agent collections and inform his own agent memory and harness design decisions.
queries asked of Scott's wikis
- agent harness subagent roster quality patterns
- context window budgeting for multi-agent systems
- tool declaration discipline as agent design invariant
- duplicate agent detection in shared rosters
- underspecified agent consequences in coding agent harnesses
- agent roster auditing methodology
2026-08-07T12:29:02Z
The roster-specific claims remain dependent on a single self-published audit, with no independent replication or renewed discussion after repeated checks. The broader context-cost mechanism survives elsewhere, but this episode has faded without advancing its core hypothesis.
2026-08-02T05:26:54Z
No independent audit has arrived, and the originating discussion is static. The broader context-cost mechanism remains plausible, but the roster-specific duplication, underspecification, and fixed-injection claims still rest on one audit.
2026-07-30T14:22:42Z
Independent Claude Code telemetry now supports the broader context-repetition cost mechanism, making the roster-overhead claim more plausible. It does not independently verify the audited roster’s duplicate, underspecified, or per-turn injection findings, so the core hypothesis remains open.
2026-07-30T14:21:34Z
evidence attached: hn.story.49110007 — Independent Claude Code telemetry supports the broader open hypothesis that repeated history and context create substantial fixed per-turn overhead.
2026-07-29T13:27:09Z
grounded: novel/high — Directly relevant to Scott's work on coding-agent harnesses, context management, and agent rosters. The audit reveals practical inefficiencies — high per-turn c
2026-07-29T13:26:11Z
origin walked (codex/luna, conf 0.98): anchor reddit.post.1v9or88 -> echo.github.d88b01951c by sshworld
2026-07-29T13:22:33Z
case created — The post presents concrete but as-yet unverified measurements of a distinct static roster-design failure not covered by the existing dynamic subagent budget-exhaustion case.