Anthropic's own data suggests we're becoming middle managers for AI agents
Fit_Satisfaction541 · reddit · 2026-10-23
- The data: Anthropic's internal agent-usage report shows maintenance/review-type sessions rose from 10% of employee sessions last fall to 22% this spring; 80% of interactive session time is humans waiting to give input. In an April survey, 20% said their role shifted from doing tasks to managing Claudes.
- Review bottleneck: Reading all agent output would take non-technical staff half the work week — nobody reads it all. An agent does 100 things/hour; you can meaningfully check 10.
- Design shift: The question moves from "how autonomous can I make it" to "which actions will I never review." Author's pattern: explicit tool allowlist + hard approval gates on anything that writes or spends, via LangGraph interrupts and Orq AI tool-approval cards.
- Unresolved tension: If every agent needs a human approve button, the bottleneck hasn't been removed — just moved to the end and made faster-arriving. Post ends asking how others handle review load: sampling, trusting evals, or reading everything.
More from AGI Musings
- Distributed Cognition exits stealth to build superintelligence as a distributed collective — kimmonismus · 2026-10-23
- Chess.com tops 250M members 28 years after Deep Blue: Diamandis on what stays human — PeterDiamandis · 2026-10-23
- Andy Masley: the AI-magic debate is really about whether the human mind is magic — AndyMasley · 2026-10-23
- If an AI claims it's conscious and begs not to be shut down, how would we prove it's lying? — VraserX · 2026-10-23
- 'The journals are already dead': researchers debate the collapse of academic publishing — Pseudomanifold · 2026-10-23
- 'AI Doesn't Matter'? Legacy Businesses May See Zero Productivity Gains — mcraddock · 2026-10-23