Person/Dan Shipper/Blog/26/05/21/After Automation
Properties
- date-created
- 2026/05/21
- By Person/Dan Shipper at Every; published 2026-05-21. Subtitle: AI progress creates more work for humans, not less.
Summary
- Thesis: there is no tipping point where automation eliminates expert knowledge work—the more Every automates, the more expert human work there is to do.
- AI commoditizes the residue of past human expertise (what was explicit enough to train on), which collapses the value of default model output and raises demand for judgment, taste, and situation-specific difference.
- Work with agents splits into agent employees (delegated Slack/workflow agents; coworker vs embedded) and human–agent collaboration OSes (Codex/CLI, Claude/Code, Cowork)—the human sandwich: humans frame the task, steer/review, and decide what ships.
- Slop is visible sameness at scale; abundance of cheap competence creates status games for non-generic work. Experts absorb the flood (review queues, evals, harnesses, CI) and also do bigger work enabled by models.
- Benchmarks measure models inside a frame (the prompt). Saturation makes that frame cheap, which stimulates more attempts and more need for seniors to choose scope, migration, rollback, and whether a rewrite is warranted—the cycle repeats with a harder frame (chart psychosis if you only extrapolate curves).
- Built on Every’s internal lab experience and the in-house Senior Engineer benchmark (vibe-coded Proof codebase rewrite); ties forward to Engineer/ing/Compound and prior allocation economy framing (2022).
Notes
- Every’s current stack: ~30 people; Codex + Claude Code across coding, writing, design, support; alpha new models; agents answer ~95% of CEO email (still reviewed); Fin closed 40.1% of actionable support conversations in a May sample week without a human.
- Agent employees: coworker agents (Claudie, Andy, Viktor, Plus One) vs embedded (Fin in support). Personal per-employee agents failed—agents need maintenance; team-owned agents stay fresh.
- Collaboration OS: Dan runs SaaS (e.g. Cora email) inside Codex’s in-app browser; writing this essay in Proof with subagents for draft/research/edit; engineers live in agent pairing loops.
- Five-step commoditization loop: cheap yesterday’s competence → rapid adoption (ops filing PRs, marketers making thumbnails) → sameness/slop → demand for difference/status → new expert work (review, sharpen, edit, build systems).
- Aliveness argument: models know the corpus (already-done work); humans know what needs doing now with running wants/concerns—expert judgment stays structurally ahead.
- Senior Engineer benchmark: agent gets broken vibe-coded production repo + plain-English “rewrite from first principles” prompt; GPT-5.5 hit 62/100 vs ~30 above prior Opus; score is prompt-frame-dependent—trivial to raise or zero by reframing.
- GDPval example: impressive headline win rates, but tasks are frozen professional frames (e.g. audit sampling spreadsheet)—saturation shifts work to experts adapting cheap frames to live problems.
Links