06-reference

every gpt56 knowledge work

2026-07-10·reference·source: Every·by Dan Shipper
gpt-5-6knowledge-workagentic-loopsai-model-comparisoneverydan-shipperchatgpt-workcompound-engineering

"How GPT-5.6 Changes Knowledge Work" — @danshipper

Why this is in the vault

Shipper's "tend the loop" framing retroactively names the architecture RDCO has been building — and his competitive read on Fable vs. GPT-5.6 is direct intelligence for Ben's phData client conversations on model selection.

⚠️ Sponsorship

Every promotes its own AI product suite (Tend, Cora, Spiral, Sparkle, Monologue, Proof) throughout the essay and footer. The Tend project — an open-source prompt + repo built on Cora — is positioned as the practical takeaway from the piece. Author is Every's CEO; this is integrated house promotion, not a paid external sponsor.

The core argument

GPT-5.6 Sol (inside ChatGPT Work, formerly Codex) crosses a reliability threshold that makes continuous knowledge-work loops practical for non-technical users. Rather than using AI task-by-task, you build a system that: (1) gathers and makes sense of information, (2) proposes decisions, (3) executes approved ones, and (4) compounds your feedback over time to do more autonomously. Your job shifts from doing the work to tending the loop — the same philosophy as compound engineering, now extended from software to knowledge work broadly.

Shipper demonstrates with email: GPT-5.6 Sol watches his inbox, researches what needs it, drafts replies, and derives preferences from his revisions run-to-run. He applies the same loop pattern to hiring, company operations, furniture shopping, editorial planning, customer research, and consulting delivery.

On model comparison: GPT-5.6 Sol wins on speed, cost, and accessibility for non-technical users. Direct quote: "Fable can do all of the above, but it's too expensive, too powerful, and too slow for non-technical users. It often speaks in its own language that even programmers have a difficult time understanding. The Claude desktop app can also do much of this, but it's hampered by hard-to-understand security controls and differences between Claude Code and Cowork's features and capabilities."

Every is releasing Tend — an open-source prompt + repo — to help readers build their own work loops with Cora, Gmail, or Slack as data sources.

Mapping against Ray Data Co

The Ray COO agent already IS this architecture — the channels agent, morning-prep, open-threads-check, check-board, process-newsletter, and Markov phase-tracker are each tended loops running on Claude on the Mac Mini. Shipper's framing arrives as a naming convention for something RDCO built before he wrote about it. The gap to close: the feedback-compounding step (loop learns from Ben's revisions run-to-run) is currently manual rather than automated — that is the next architectural tier worth designing explicitly.

Competitive signal for phData DSA work: Shipper's specific claim that Fable is "too expensive, too powerful, and too slow for non-technical users" and that Claude has "hard-to-understand security controls" maps directly to the client objection space Ben will navigate when recommending the Anthropic stack. The counter-argument is that Claude Code and Sonnet's permission model is a feature for enterprise risk controls, not a UX bug — worth pre-loading before discovery conversations.

Sanity Check angle: this essay is a model for how to cover model releases — behavioral change plus job-function reframe, not benchmark comparison. "Your job title stays the same; your job description changes" is a more defensible frame than MMLU scores, and more aligned with what Sanity Check's CFO/operator audience cares about.

Related