# Automation Relocates Toil Into Supervision ## Claim When a team routes its core work through autonomous agents, the human job shifts from producing to supervising, and the supervision overhead can exceed the toil the automation removed. <!-- atomicity: 1 idea ✓ --> ## Evidence **Source**: [[3. Resources/Readwise/Articles/Agentrification and the Agentrification Index]] The accounting error is treating removed production time as saved time. It is only saved if the work that replaces it is cheaper, and supervision has a poor cost profile: it is interrupt-driven, it cannot be batched well, it requires the same domain knowledge as producing, and it scales with the *number* of agent outputs rather than with their difficulty. This is a specific instance of a much older result. Bainbridge's "ironies of automation" observed that automating the easy parts of a task leaves humans with the hard residue plus the new burden of monitoring — and degrades the very skills monitoring requires, because the operator no longer practises them. Agentic development reproduces the pattern almost exactly, with the additional wrinkle that agent output is fluent enough to make supervision harder than it looks. ## Scope & conditions - **Holds** for routine, well-specified tasks where review and prompting cost is high relative to the work itself — which is most of what agents are currently pointed at. - **Breaks** where agents handle genuinely novel generation no human could match, since there is no prior toil to compare against and supervision buys capability rather than replacing labour. - **Breaks** where verification is automatic. If a test suite adjudicates, supervision cost collapses toward zero and the arithmetic reverses. - **Weakens** with experience: teams that build good specification and review harnesses reduce supervision cost substantially, which is why the harness matters more than the model. - **Worth measuring rather than assuming** — the claim is that overhead *can* exceed the saving, not that it always does. ## Links - [[3. Resources/Permanent Notes/Fluency Makes Small Errors Harder to Catch Than Large Ones]] — because extends: supervision is expensive partly because the artifacts being supervised are uniformly plausible, so the reviewer cannot triage by surface. - [[3. Resources/Permanent Notes/AI Compresses the Knowledge-Worker Pipeline From Both Ends]] — because extends: if the machine takes capture and first-pass synthesis, what remains at the human end is framing and judgment — which is exactly supervision. - [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] — because applies-in: the checker is the supervision, and its cost is the thing this note says goes unbudgeted. - [[3. Resources/Permanent Notes/Maintainer Attention Is the Scarce Resource in Open Source]] — because extends: the same relocation at ecosystem scale, where generated contributions convert contributor effort into maintainer review load. - [[3. Resources/Permanent Notes/Rewarding AI Usage Manufactures Usage Not Value]] — because contrasts: measuring adoption counts the production side while leaving the supervision cost invisible, which is how the saving gets overstated. - [[3. Resources/Permanent Notes/A Declared Capability Costs You Every Turn Not Every Use]] — because extends: both are costs that sit outside the obvious ledger and only appear when someone measures. - [[AI & the Future of Knowledge Work MOC]] — because applies-in. ## Implications - Any productivity claim for agentic workflows needs a before/after that counts *review and prompting* time, not just production time. Most published claims do not. - It argues for investing in verification harnesses over model upgrades — automatic adjudication is the only thing that actually removes the supervision burden rather than moving it. - For managers, the failure mode is invisible: headcount-equivalent savings appear on the production side while the supervision cost lands as diffuse senior-engineer interruption that nobody books. ## Tests / falsifiers `proof_state: working` - A matched before/after time-and-motion study on comparable tasks, counting review and prompting time as work rather than overhead. If total time falls and quality holds, the claim is wrong for that class of task. - If teams with mature test harnesses show no supervision premium, the scope condition above is doing all the work and the claim is narrower than stated. ## Open Questions - Does supervision cost fall with team experience, or does it plateau because the irreducible part is judgment? - Is there a task-complexity threshold above which supervision becomes cheaper than production, and can it be identified in advance? - Does Bainbridge's skill-degradation half apply — do supervising engineers lose the ability to produce, and how fast? ## Source - [[3. Resources/Readwise/Articles/Agentrification and the Agentrification Index]] — the supervision-overhead claim and the agentrification framing