Anthropic's own data suggests we're becoming middle managers for AI agents
retained text
anthropic published data on what agents did to work inside their own company this month, and this post isn't about 6-to-1 hours but about how we are spending our time on and it's mostly maintainance and task like building skills for agents, reviewing their output, babysitting pipelines, went from 10% of the average employee’s agent sessions last fall to 22% this spring which is about 80% of interactive session time is the human sitting there waiting to give input and in an april survey, 20% of people said their role had shifted from doing the task to managing claudes and reading everything their agents produce would take even non-technical staff half their work week, so nobody reads it all that last stat is the whole thing for me like once you can’t read all the output and deciding which 5% a human has to look at before it ships you’re managing something that types faster than you can review which is exactly the problem everyone building agents hits the moment they give one real tool access like the agent can do 100 things an hour and you can meaningfully check maybe 10. so the design question shifts from how autonomous can i make it to which actions am i willing to never review, and which ones have to stop and wait for me, okay like reading a doc, sure, let it rip sending the email, moving money, deleting the row, those wait the pattern that’s held up for me is an explicit allowlist of tools plus a hard approval gate on anything that writes or spends, I use tools for specific purpose like langgraph to the gate with interrupts and Orq AI to flag individual tools as requiring approval so the agent pauses on a tool approval card until a human signs off. either way the agent’s speed stops mattering for the actions that can’t be undone, because those move at human speed on purpose the part i can’t resolve is that this doesn’t scale the way the automation was supposed to. if every agent needs a human on the approve button for the actions that matter, and reviewing is already eating half the week, then the human bottleneck hasn’t been removed, it’s been moved to the end and made faster-arriving how are you handling the review load? sampling a percentage, trusting evals instead of reading, or actually reading everything that clears a threshold
document fetched 2026-10-23 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xcqkgw/anthropics_own_data_suggests_were_becoming_middle