06-reference

every still employed when ai does everything

2026-06-04·reference·source: Every·by Laura Entis
ai-laborlump-of-laboragent-orchestrationl5-unhobblingcustom-skillsenterprise-roadmaps

"Why We'll Still Be Employed When AI Can Do Everything" — Laura Entis (Context Window issue; lead Counterpoint by Mike Taylor)

⚠️ Sponsorship

House promo, not third-party. The issue opens with a Launch block for Spiral 4.0, which is Every's own product (sponsor_entity: self). The push: new "style engine" that drafts in your voice, a new MCP alongside the existing CLI/API, expanded team workspaces, and a switch from session limits to token-based pricing (personal plans drop $25 → $15/mo, team $35 → $25/user). The footer also cross-promotes Every's full product line (Sparkle, Cora, Monologue, Proof). No external paid sponsor in this issue — sponsorships@every.to is solicited but unfilled here. Flagged so the Spiral framing is read as self-interested, not neutral.

Why this is in the vault

Two reasons. First, the Counterpoint is a named, public intra-company rebuttal of Dan Shipper's "After Automation" thesis — Mike Taylor argues the "humans always stay one frame ahead, forever" claim breaks, but we stay employed anyway on a cost/compute argument rather than a capability one. That's a sharper, more falsifiable version of the lump-of-labor debate RDCO has been tracking, and it directly maps to the COO-agent economics question (when is it worth the tokens to have Ray do X vs. a human). Second, the Steal This Workflow item is a clean, concrete articulation of the "build the most focused local skill, don't download generic ones" discipline — which is exactly RDCO's ~/.claude/skills/ philosophy and the harness/orchestration lane.

Issue contents

Five-segment hybrid (Context Window daily format, curated by Laura Entis):

  1. Launch — Spiral 4.0 (house promo, see Sponsorship). Voice-drafting tool goes agent-native with an MCP; token-based pricing.
  2. Signal — "Enterprise AI product roadmaps are hard" (Mike Taylor). Microsoft moved fast on OpenClaw (Nadella's "virus" security framing in Nov 2025 → internal "ClawPilot"/Project Lobster by May → "Scout" personal agent shipped at Build), and still shipped behind the hype cycle — OpenClaw search interest spiked in January and decayed within months. Lesson: months-ahead keynote planning can't keep pace with a news cycle measured in weeks; "maybe next time we'll just announce it on X." (Note: "OpenClaw," "Opus 4.5/6," "GPT-5.5/7," "Hermes," "Gemini Spark" are the issue's near-future/illustrative product names.)
  3. Steal This Workflow — "Make your agent more efficient with custom skills" (Naveen Naidu, Monologue GM). Three steps: (a) ask the agent "what tools can I give you to work faster?"; (b) build the most focused local skill possible — Naveen had Codex write a small repo-local CLI script hitting Fin's API to pull the active support conversation into markdown, rather than relying on Fin's generic 13-action MCP; (c) tell the agent when to use it via a project-level instruction. Rule of thumb (paraphrased): don't download skills — interact, find the inefficiency, then have the agent author the skill.
  4. Counterpoint — "AI will outpace human ability, but it won't be cheap" (Mike Taylor, "—MT"). The lead/subject piece. See core argument below.
  5. One Last Thing. Light segment on model verbal tics the Every team can't get models to drop (Claude's "locked in"/"load bearing," Codex's "my instinct is," "leave it here and pick it up in the morning").

The core argument

The Counterpoint (Mike Taylor) responds to Dan Shipper's "After Automation," which holds that AI progress creates more human work forever: each time models saturate a benchmark and make yesterday's human competence cheap, humans "reset the frame" to higher-level work, and "the frame is never the framer."

Taylor concedes the dynamic matches his lived experience (prompt-writer → context-supplier → agent-orchestrator, frame expanding each time) but rejects the forever. His prediction: within a year or two, in a few well-run companies, AI executes every knowledge-worker task better than humans — including setting frames. Framing isn't magic; goals are derived from embodied experience and physical constraints ("physics is the ultimate eval"), and a system that learns from its environment can run the same loop.

The twist is the economic save, not a capability one. Intelligence costs energy; evolution already optimized humans to be cheap-but-good via heuristics/intuition, while a model must brute-force the same answers with expensive "thinking tokens." So the operative question shifts from "Can AI do this?" to "Is it worth the compute?" Past some price (he gestures at a hypothetical $2,000/mo or $20,000/mo model), you'd rather point the expensive model at curing cancer than at making slide decks — so you just hire the human. Evidence cited: Waymo is a safer driver than humans and charges a fraction of Uber/Lyft, yet the city rideshare workforce grew. Conclusion: models outpace us in raw capability, but we stay employed anyway because some buyers prefer human work, especially when it's cheaper.

Mapping against Ray Data Co

Strong. Two distinct lanes hit.

Counterpoint → L5 / agent-deployer lane + the lump-of-labor framing. This is the most useful piece for RDCO. It reframes the "will the COO agent replace the human" question as a unit-economics question, not a capability one — which is exactly the lens RDCO should hold while building toward L5. The "is it worth the compute?" pivot is the operational version of [[project_l5_north_star_strategic_direction]]: the bet isn't that Ray can do everything, it's that Ray is cheap enough to do the high-frequency, lower-stakes knowledge work while the founder's (expensive, scarce) judgment stays pointed at the frame-setting that's actually worth it. The Waymo example is a direct lump-of-labor data point — capability arrived, workforce grew anyway — corroborating the demand-side thesis RDCO has been tracking (see [[2026-06-04-lassie-smb-ai-frontier-steijn-pelle-assessment]], where the whole SMB-agent demand case rests on lump-of-labor being false). This is also the rebuttal to [[2026-05-27-every-after-after-automation]] — read the pair together: Shipper says humans stay ahead forever; Taylor says they don't, but stay employed on cost. RDCO's position is closer to Taylor's — don't bet the company on staying one frame ahead of the models; bet on being the cheaper-than-the-frontier-model deployer of agent labor.

Worth a caution flag per [[feedback_calibrate_overconfidence]]: Taylor's "AI sets frames better than humans in 1-2 years" is an aggressive claim, and his cost-floor argument leans on the assumption that thinking-token cost won't collapse — if inference cost craters (the trend), the "just hire the human" floor erodes and the comfort is thinner than it reads.

Steal This Workflow → harness / orchestration lane. Naveen's three-step "build the most focused local skill" pattern is a direct restatement of RDCO's skills-over-commands discipline ([[feedback_skills_over_commands]]) and the dynamic-harness thinking in [[2026-06-02-thariq-dynamic-workflows-harness-for-every-task]]. His "don't download generic skills; find the inefficiency, then have the agent author a narrow one" rule is a clean external validation of the thin-harness/fat-skills bias — narrow repo-local script beat the vendor's generic 13-action MCP. Concrete take-away for the RDCO skill library: when a recurring task is slow, the move is to ask Ray to scaffold a focused skill for that exact task, not reach for a broad pre-built integration.

Signal (enterprise roadmaps) → weak/contextual. Useful as a reminder that ship-cadence in the agent space is measured in weeks, not the months an enterprise keynote cycle assumes — relevant to how fast RDCO's own tooling decisions should turn over, but not load-bearing.

Related