Notes — Dan Shipper on Every, AI-Native Company Building, and the Allocation Economy

Four questions [Adler frame]

Q1 — What is it about?
Every as an AI-native institution (newsletter + SaaS + consulting), and the two frameworks Dan has developed from running it daily: (1) the allocation economy thesis — management skills become universal as AI does knowledge work; (2) compounding engineering — building systems where each unit of work makes the next cheaper.

Q2 — How is it argued?
From lived practice, not theory: Dan runs a 15-person company with no manual code-writing, built a product (Cora) for $300K that wasn’t technically possible three years ago, and consults companies on AI adoption. The frameworks are inductions from this operational experience. Allocation economy article was written ~2.5 years ago (pre-agents) — it was a forward projection from early ChatGPT use patterns.

Q3 — Is it true?
Allocation economy thesis: compelling directionally. The management-skills-become-universal claim is plausible but may underestimate friction — management is partly learned by managing humans, and it is unclear whether managing AI agents develops the same skills. The 8% → broad distribution claim is a long-run bet, not near-term.
CEO as adoption predictor [?]: anecdotal. Plausible and consistent with other sources (Chip Huyen, Bret Taylor on organisational change), but not measured.
AGI definition as economic threshold: more operational than scientific, but useful for distinguishing near-term agentic automation from the stronger claim.

Q4 — What of it?
Allocation skills (evaluating quality, scoping, feedback, vision) are worth building now — they are the durable meta-skill regardless of whether the allocation economy thesis is precisely right. Compounding engineering is immediately applicable: encode each feedback session as a prompt; build automations rather than repeating manual processes.

Glossary

Allocation economy — Dan’s term for a post-knowledge-economy in which humans manage AI agents rather than doing tasks directly. Management skills (currently rare) become the scarce, broadly demanded capability.

Compounding engineering — designing each unit of work to make the next unit easier. The engineering of the engineering system. Coined at Every by Kieran and Nityesh. See Compounding Engineering.

Head of AI operations — role responsible for translating the principal’s taste into reusable prompts, automations, and knowledge bases. Not a developer role. Dan’s prediction for an emerging standard job.

AGI (Dan’s operational definition) — the point at which it is profitable to run agents indefinitely without human oversight. Economic threshold, not capability threshold. Winnicott leash analogy: the leash disappears when the economics of oversight change, not when capability crosses a bar.

Sip seed — Dan’s funding structure: capital committed but drawn down on demand. Preserves optionality and prevents the psychological burn-it-all effect of a large bank balance.

Context engineering — getting the right context to the model at the right time. Dan had called this “knowledge orchestration” for three years before learning Tobi Lütke coined the current term. [§ Compounding engineering section]

Cora — Every’s AI email assistant. Built for ~$300K total (salaries included), by 2 engineers and 15 Claude Code instances. Reference case for agentic engineering cost collapse.

GPT wrapper — pejorative term for AI products that wrap a foundation model API. Dan argues these are legitimate products, unduly maligned; the specialisation and UX is the product.

Every’s model

Every has 15 people across four lines: newsletter, SaaS products (Cora, Spiral), consulting, and incubation. Nobody writes code manually. The model is described as a “creative playground” and institution simultaneously — held in tension by keeping fundraising small.

Key data point: Cora (email AI) built for approximately $300K all-in including salaries. Product was not technically possible three years ago at any budget. Demonstrates the cost collapse in AI-native product development.

Model stack as of this episode: o3 for personal/writing tasks; Claude Opus for writing judgment (“first model that can truly judge whether writing is good” [?]); Claude Code for all engineering; Gemini for app internals/cost-sensitive routes; Codex for one-off features.

Compounding engineering in practice

The mechanism at Every:

Every feedback session with a writer is recorded and converted into a prompt.
Prompts accumulate in a shared library accessible to all agents and team members.
Head of AI operations (Katie Parrott) is the role that converts tacit CEO judgment into these machine-readable assets.
Alex Duffy case: young writer recorded all of Dan’s feedback as prompts. Made a year’s progress in two months.

The compounding effect: each iteration starts from a higher baseline. The key investment is not in the individual task but in the encoding infrastructure. [§ Head of AI operations]

Allocation economy

The article was written ~2.5 years before this episode, when agents were not yet considered viable. Dan inferred the thesis from early GPT-3/4 usage patterns.

The observation: when using early models, Dan found himself spending most of his time on:

Communicating the problem clearly
Gathering the right context
Selecting which model to use
Dividing up tasks across models
Evaluating output quality
Giving feedback
Holding a vision and success criteria

These are precisely the activities of a manager. The conclusion: AI shifts the scarce skill from doing to allocating.

Skills that become more valuable per Dan (also noted by Lenny from the article):

Evaluating talent/output quality
Vision (knowing what to want)
Taste
Knowing when to step in vs. delegate

Management skills were rare because managing humans was expensive and hard. Managing AI agents is cheap. Therefore these skills will spread. The distribution will shift. [§ Allocation economy]

CEO adoption predictor

Dan’s consulting finding: the number one predictor of org-wide AI adoption is daily CEO use of ChatGPT/Claude. Two mechanisms:

Excitement transmission. CEO enthusiasm spreads.
Expectation calibration. CEOs without personal intuition for the tools either generate no adoption or impossible expectations.

Walleye hedge fund ($10B AUM) as reference case: wrote the AI-first memo with ChatGPT (announced in the memo); weekly prompt-sharing meeting; weekly usage stats email with early adopter recognition. Result: org-wide momentum.

The 10/80/10 rule from consulting experience: ~10% early adopters, ~10% resistant, ~80% will adopt if given exact prompts for their specific job. The consulting method: customise training per team based on pre-interviews.

AGI as economic threshold

Winnicott’s leash in developmental psychology: parents externalise their values through supervision. When the child has internalised those values, the leash is no longer needed — not because capability threshold is crossed, but because the economics of supervision change.

Dan applies this to AI: AGI is when the economics of running an agent indefinitely without human oversight become profitable. This is operationally useful: it shifts the question from “can the model pass a test?” to “does the math on running it unsupervised work out?” [§ AGI definition]

The “profitable infinite agents” formulation distinguishes near-term agentic work (still requires oversight, still has cost) from a qualitative transition where oversight is no longer economically warranted.

Writing and generalism

Generalism at Every: everyone is a generalist because AI handles the specialised sub-tasks. Dan argues AI may reverse the specialisation trend that civilisation imposed (Athens generalisation → empire specialisation → AI re-generalisation).

Dan’s writing arc: stopped writing early at Every to “run the company.” Growth stalled; Dan was unhappy. Returned to writing as the centre. Realised writing-centric company patterns exist (Joel Spolsky, Jason Fried, Bill Simmons, Sam Harris). The founder IS the product in a media business.

Key contrast with tech startup pattern: in a startup, you build v1 and hire people to build the rest. In a media business, if the founder stops making the product, you lose product-market fit.