From Doing to Tending — Every Context Window (2026-07-12)

Canonical URL: https://every.to/context-window/from-doing-to-tending

Why this is in the vault

The issue crystallizes the "tending vs. doing" mental model for AI-augmented knowledge work — the idea that as models become capable of running full loops, the human role shifts from executing tasks to tending the systems that execute them. That framing is directly load-bearing for how RDCO is designing the COO agent architecture. The Grok 4.5 benchmarks provide current pricing/quality data for the frontier (Grok at $2/$6 vs. Claude Opus 4.8 at $5/$25 per M tokens). Ashwin Sharma's Alignment column on AI scribes and cognitive deskilling is a precise, clinically-grounded articulation of the deskilling risk that underlies Ray's verification/fresh-eyes SOPs.

Issue contents

Main essay (subscriber-gated, click-through): "From Doing to Tending" — not reproduced in the email body. Theme: the shift from doing knowledge work yourself to tending AI loops that run the work. The Dan Shipper piece below is the intellectual scaffolding.

Mini-Vibe Check: Grok 4.5 Is Fast, Cheap, and Finally Useful (by Katie Parrott)

• Jointly trained by Cursor and SpaceXAI on codebase interaction data
• Every team consensus: Opus-level. Mike Taylor benchmarks it slightly above Claude Opus 4.8 on a multi-step assignment (Grok followed every step + returned complete result; Opus stopped early or skipped parts)
• Kieran Klaassen ran it through /LFG workflow — puts it at "Claude Opus 4.5-to-4.6 range. Not state of the art, but pretty good for lots of things, and very fast"
• Speed: ~80 tokens/sec, ~2x token efficiency of leading models
• Pricing: $2/$6 per M input/output tokens vs. Claude Opus 4.8 at $5/$25, GPT-5.6 Sol at $5/$30
• Holds its own on vibe coding; still loses to Sol on writing quality
• Verdict: "You probably don't need to swap out your daily driver. But if you already work in Cursor, Grok 4.5 is right there, and it's earned a slot for long, multi-step assignments where speed, price, and follow-through count more than the last few points of quality"

Knowledge Base curation (5 articles, summaries only — editorial, not sponsored):

1. GPT-5.6 Sol Is Our Favorite Model to Collaborate With (Katie Parrott / Vibe Check) — Sol as the team's daily driver for everything except the biggest/loosest assignments (those go to Fable)
2. How GPT-5.6 Changes Knowledge Work (Dan Shipper / Chain of Thought) — introduces "Tend," an open-source prompt/repo for building loops that run knowledge work; "you tend the loop instead of doing the work yourself" — the conceptual heart of this issue's title
3. Use Fable Before You Know What to Ask (Katie Parrott / Context Window) — Fable earns its premium on jobs where the goal itself is still unsettled
4. Welcome to Efficiencymaxxing (Laura Entis / Context Window) — "revenue per million tokens" as successor metric to revenue per employee
5. How a Writer Uses AI Without Losing His Voice (Dan Shipper / AI & I) — Craig Mod keeps a WiFi-free MacBook for writing; vibe codes SaaS replacements; AI raises the premium on human intent

Alignment column (Ashwin Sharma — full text in email):

• Argument: clinical writing is not documentation overhead — it is the cognitive act that produces clinical reasoning. AI scribes offload that act, which atrophies judgment.
• Aviation parallel: autopilot made flight safer, but mandated simulation training was required to preserve manual skills for when automation fails or hands back control.
• Predicted arc: "First, we'll automate too much — because it reduces burnout and documentation misery. Then we'll panic about skill loss. Eventually, I hope, we'll land somewhere sensible: AI scribes for routine cases, manual reps and simulation mandated for students, young doctors, and anything complex."
• Author note: Sharma is a former junior doctor now in health tech.

Events/products: Every IRL July 15 (Brooklyn, paid subscribers); Monologue now has web editing at app.monologue.to.

Mapping against Ray Data Co

The "tending vs. doing" frame is the clearest public articulation of where RDCO is in its L4→L5 build. The COO agent is being architected to run loops (gather info → propose decisions → execute approved ones) while Ray handles judgment calls — Dan Shipper's "Tend" repo is the open-source analogue to what Ray is building proprietary. This issue's theme directly validates the direction; the Tend framework is worth pulling and examining for prompt patterns.

The Grok 4.5 pricing data ($2/$6 per M) is actionable: for long, multi-step tasks inside the agent stack where speed and cost matter more than top-1% quality, Grok 4.5 inside Cursor is a credible cheaper-tier option. The benchmark result (Grok above Claude Opus 4.8 on follow-through on a complete multi-step assignment) is noteworthy given that Opus 4.8 is what RDCO currently uses for heavy reasoning tasks.

The Ashwin Sharma Alignment column is the verification/fresh-eyes SOP stated in medical terms. The "cognitive offloading" risk — where the muscle atrophies because you stop exercising it — is the exact failure mode Ray's fresh-eyes subagent gate is designed to counter. The aviation model (mandated simulation training to stay sharp for when automation fails) is a good framing device for how to think about when to run manual verification passes vs. trusting the agent chain. Worth linking if a Sanity Check piece ever tackles AI deskilling.

Grok 4.5 pricing note: at $2/$6 vs. Sonnet-class at higher prices, this is the first xAI model with real cost arbitrage for production batch workloads.

Related

• [[06-reference/2026-07-10-every-gpt56-knowledge-work]] — Dan Shipper's "How GPT-5.6 Changes Knowledge Work" introduces the Tend framework and "tending the loop" concept that this issue's title is built on; the Knowledge Base entry in this issue points directly back to it
• [[06-reference/2026-05-24-every-cheap-competence-new-frontier]] — Ashwin Sharma's prior Alignment column on AI and medical situated judgment; the Alignment section in this issue is a direct sequel (same author, same clinical-AI theme, extends the argument to skill atrophy specifically)