concept · concepts/The Human Sandwich.md

The Human Sandwich

status: active · created 2026-06-02 · updated 2026-11-08 · 2 sources · #agents #human-in-the-loop #workflow #collaboration

The work pattern for complex agent collaboration, coined by an Every employee: humans are the bread on either side of the AI's work — a human sets the frame, the AI collapses the task, and a human judges and extends the result.

The pattern

For any complex, original task, Dan Shipper reports the best results come from "an AI and a human going back and forth in the same workspace" (Codex, Claude Code, "Claude Cowork") — not from handing work off and walking away. The three beats:

Human sets the frame — what we're trying to do and what counts as good.
AI collapses the task — drafts, searches, codes, summarizes, compares.
Human judges and extends — Is this good? Where does it belong? What should happen next?

The human owns the start and end of each agent task: defining the frame, verifying the output is done well, and finding the next piece of work. This is why, per Shipper, "both of these modes require you or another human in order to work well." The sandwich is the concrete shape of the collaboration mode (as opposed to fully delegated "employee" agents).

Why it matters

The human's contribution sits exactly where models are weak: judgment about what needs to be done right now and what counts as different rather than default (AI Commoditizes Expertise). The sandwich is the operational reason agents "still need humans."

The sandwich in a wet lab (2026)

Anthropic — Investigating Vault Particles with Claude is the pattern outside knowledge work. The team's own verbs: Claude worked "under our direction," they "advised on the analysis decisions," and when the cryo-EM processing went off-script they "told it to stop improvising" — "Please let's do normal things unless they don't work." Humans chose the 44 candidate proteins (frame) and judged the resulting maps (extend); Claude searched 15 billion sequences, designed constructs and ran RELION (collapse).

Two refinements the case adds:

The bread has to be load-bearing. The team's errors were caught fast where they had expertise and ran for two weeks where they didn't. The judging slice only works if the judge knows the domain (Verification Requires Domain Knowledge).
The frame step includes overruling the model. Claude argued against the purification method that ended up working. A human who only approves or rejects what the model proposes would have lost the 48× speedup.

Connections

Verification Requires Domain Knowledge — the precondition for the "judge" slice.
Agents as Employees vs Human-Agent Collaboration — the sandwich is the signature pattern of the collaboration mode.
AI Commoditizes Expertise — the "judge and extend" slice is where human difference defeats Slop.
The Infinite Backlog — the human's "what's next?" step is how the backlog gets prioritized.
Agent Autonomy Spectrum — the sandwich implies semi-synchronous working, not full autonomy.
The Bottleneck Shift — the human "frame" and "what's next?" steps are exactly the what-to-do judgment Ethan Mollick calls the new bottleneck.
Communication and Control Are Intertwined — see also: framing + steering is control exercised through communication.
Human-in-the-Loop by Exception — cross-domain twin: the data-pipeline version of the same pattern — automate the collapsible bulk, reserve the human for the judgment slice (and bank that judgment so it's paid once).
Placement and Judgment Over Raw Execution — the sandwich is the AI-side instance of Move 2 ("reserve human judgment for the exception") in that synthesis.

Sources

The AI Daily Brief — Why Agents Still Need Humans — via Dan Shipper — After Automation.
Anthropic — Investigating Vault Particles with Claude — the pattern observed in a biology lab, including its failure mode when the human lacks the domain.

Linked from 23

Agent Autonomy Spectrum

(The Human Sandwich lives here.)

Agents as Employees vs Human-Agent Collaboration

Its signature shape is The Human Sandwich.

AI Commoditizes Expertise

The Human Sandwich — the "judge & extend" step is where humans inject difference.

Anthropic — Investigating Vault Particles with Claudeauto-ingested

The Human Sandwich — "under our direction," "we advised on the analysis decisions," "we told it to stop improvising" is the pattern in a wet lab.

Claude Code

The Human Sandwich — the collaboration pattern it enables.

Codex

The Human Sandwich — the collaboration pattern these harnesses enable.

Communication and Control Are Intertwined

The Human Sandwich — see also: a human steers an agent by framing and judging — communication as the control surface.

Dan Shipper

Champions the human sandwich pattern (coined by one of his employees).

Dan Shipper — After Automationto read

NLW read long excerpts but explicitly skipped sections — notably Shipper's full treatment of the AGI objection ("doesn't better AI just do the expert work too?"), described as "an interesting meditation on the nature of benchmarks." Getting the primary sour…

Every

The Human Sandwich — pattern coined by an Every employee.

Front Door Pattern Language

Hierarchy, most-linked, measured 2026-11-08: Method Reflection and The AI Daily Brief — Why Agents Still Need Humans (29 inbound each), Working Heuristics — Data, Matching, Product, Org, Tooling (26), The Human Sandwich (24), The Bottleneck Shift (23).

Human-in-the-Loop by Exception

It's the data-pipeline cousin of The Human Sandwich.

Overview

AI and work (the founding domain) — The Bottleneck Shift (deciding what to do is the new scarce input), AI Commoditizes Expertise → Slop → demand for difference, The Human Sandwich and the two modes of working with agents, Personal vs Team Agents.

Placement and Judgment Over Raw Execution

AI & work: The Human Sandwich — the human frames and judges; the agent collapses the middle.

Provable Inferenceauto-ingested

The Human Sandwich · Human-in-the-Loop by Exception — human review is one kind of verifier; PI shows what it looks like when the verifier is a small trusted program instead.

Slop

The Human Sandwich — the human "judge & extend" step is the anti-slop checkpoint.

The AI Daily Brief — Why Agents Still Need Humans

Shipper considers the second mode the more important; its signature pattern is The Human Sandwich — human frames the task → AI collapses it (drafts, search, code, summarize) → human judges and extends.

The Bottleneck Shift

The same scarcity also drives The Infinite Backlog (the human's job becomes choosing what's next for the agent) and The Human Sandwich (the human "frame" and "what should happen next?" steps).

The Infinite Backlog

The Human Sandwich — one answer to the backlog: humans manage the start and end of each agent task, choosing "the next piece of work to do."

Verification Requires Domain Knowledgeauto-ingested

Human-in-the-Loop by Exception and The Human Sandwich both reserve judgment for a human.

Vibe Coding

The Human Sandwich — vibe coding is human-frames / AI-executes in the coding domain.

What's Scarce When Execution Is Cheap

The Human Sandwich — the operational form: the human supplies the framing and the judgment AI can't.

Working Heuristics — Data, Matching, Product, Org, Tooling

The Human Sandwich — the matching trio (automate the bulk / persist decisions / make the manual step teach) is the data-pipeline cousin of the human-in-the-loop sandwich; both reserve the human for the judgment slice and bank the rest.

Links to 18

In the log 3

2026-11-08 auto-ingest

Anthropic — Investigating Vault Particles with Claude (Anthropic Life Sciences, 2026-11-04; subscribed-feed capture) — unattended; 1 source + 1 concept (Verification Requires Domain Knowledge — you can only catch AI errors where you already have expertise), both tagged auto-ingested. Deliberately lean per T5 salience. The wiki's first worked case of AI-accelerated science: prokaryotic vault structure (phage L05, 96 copies, 9.5 MDa) in <4 months by a team outside its expertise; purification 3 days → 90 min, first model in 34 h; Claude's errors documented (2 weeks lost to a bad protocol; mirror-image cryo-EM map; "we told it to stop improvising"). Updated AI-Accelerated Science (case study + 1 contradiction: physical bench time treated as the cap, but a method substitution cut a physical step ~48× with no lab automation), AI Use Tracks AI Capability (verifiability has two halves — ground truth exists vs operator can read it), The Bottleneck Shift (Claude supplied method ideation + a coenzyme-A hypothesis; also argued against the method that worked), The Human Sandwich (the pattern in a wet lab; the bread must be load-bearing), Agent Autonomy Spectrum (internal multi-agent harness; the one unsupervised improvisation), AI Commoditizes Expertise (borrowed expertise and its bill), Critical Thinking Is Domain-Bound (first field evidence on its open question), Human-in-the-Loop by Exception + Falsification Is Asymmetric (see-also, no source-count bump), Anthropic (in-house biology lab), What's Scarce When Execution Is Cheap (narrows the scarce contribution from judgment toward verification), Overview, index. Method Reflection: T3 sharpened again — review only catches what the reviewer knows, which lands directly on T4's PR-review checkpoint; 4 candidate counterweights recorded. Seed Match Process Weight to Stage and Stakes: third promotion-candidate callout (verdict: analogy only — fourth near-miss; needs a direct source). 3 needs-your-call (1 contradiction, 1 promotion candidate, 1 method question). qmd search still broken (Node ABI) — searched with grep/index.
2026-06-26 query

filed synthesis Placement and Judgment Over Raw Execution (cross-domain: AI-and-work ⇄ engineering-craft share 3 recurring moves) + new hub concept Do It Once, at the Right Place (meta-principle behind Data Contract/Data Gravity/Human-in-the-Loop/freeze-rewrite). Back-linked into What's Scarce When Execution Is Cheap (companion synthesis), The Bottleneck Shift, The Human Sandwich, Personal vs Team Agents; Overview updated.
2026-06-26 ingest

Working Heuristics — Data, Matching, Product, Org, Tooling (~30 distilled heuristics, human-supplied inline → persisted to raw/ as the human's own content, NOT agent-fetched) — 1 source, 1 entity (Medallion Architecture), 14 concepts opening the new engineering-craft domain (data modeling/pipelines, matching/identity, product/platform, org/collaboration, tickets/tooling); back-linked into 4 existing pages (The Human Sandwich, Agents as Employees vs Human-Agent Collaboration, Personal vs Team Agents, Control-Flow Integrity); updated Overview (filled the empty MOC, added domain map). PROMOTE evaluated on seed Match Process Weight to Stage and Stakes → NOT promoted (ticket/tooling-hygiene craft is adjacent but doesn't assert the weight ∝ stage×stakes thesis); cross-linked instead. Cross-domain rhymes flagged: human-in-the-loop ⇄ The Human Sandwich; allowlist ⇄ Control-Flow Integrity (default-deny); build-with-not-against ⇄ Personal vs Team Agents.