Your automation isn’t set and forget, every agent needs a human on both ends
Idea
Agents don’t run themselves: a human has to frame the task at the start and judge the output at the end, so the human is the “bread” on either side of the AI’s work.
Notes & Verified Quotes
Curation, relay-an-idea. Source: Dan Shipper (cofounder/CEO of Every), “After Automation.” The framing image is Kieran Klaassen’s “human sandwich.”
The shape of the work:
QUOTE: “we’re the bread on either end of the AI’s work” — src: scratch/sources/after-automation.txt | source: https://every.to/p/after-automation
QUOTE: “a human sets the frame, AI collapses the task, and a human judges and extends the result” — src: scratch/sources/after-automation.txt | source: https://every.to/p/after-automation
Why “set and forget” fails — the agent’s quality is tied to the human attached to it:
QUOTE: “the agent needs a human in order for the work to, well, work” — src: scratch/sources/after-automation.txt | source: https://every.to/p/after-automation
Evidence it’s not hypothetical: Every gave every employee a personal agent, then pulled back because unattended agents rot.
QUOTE: “personal agents quickly get stale when the employees they were working with gave up on them” — src: scratch/sources/after-automation.txt | source: https://every.to/p/after-automation
My angle (to develop): “set it and forget it” is the selling line for most automation, and it’s a trap. The win isn’t an agent you forget; it’s a workflow you can step into at the two points that matter (framing, judging) and ignore in the middle.
Outline
Draft — LinkedIn (master)
Draft — X (variant)
Final
Related
- Three Modes of Working With AI — the modes differ exactly by how much human is on each end of the agent’s work
- Granularity - Tools Are Atomic Primitives, Features Are Outcomes — agents act well when a human frames them with the right primitives; bad framing is bad output
- Why Agent-Native Software Works Now — agent-native tools are built precisely so a human can step into the loop at start and end
- @Dan Shipper — primary source (Every); the “human sandwich” comes from Kieran Klaassen’s team
- Build a Claude skill that rewrites itself when it’s wrong — a concrete case of the trap: a skill that auto-updates in the background still needs a human to judge the “fix”
- Agents Have Autonomy, Not Agency — why the front slice of bread can’t be removed: the agent has no ends of its own, so a human must supply the wanting
- The Orchestration Tax - Cognitive Bandwidth Doesn’t Parallelize — the sandwich multiplied: every extra parallel agent adds two more slices of human framing and judging, and those don’t scale
- A Plan Document Can’t Hold Your Product Vision — why the front slice of bread can’t be compressed into a spec: judgment hasn’t been formed yet and no document can carry it