Dan Shipper on Every

Guest: Dan Shipper — co-founder and CEO of Every; essayist; AI researcher; founder of the AI and I podcast.
Host: Lenny Rachitsky
Source: Lenny’s Podcast.

Overview

Dan Shipper runs Every: an AI-native company incubator (newsletter + SaaS products + consulting arm) with 15 people and no manual code-writing. The conversation covers Every’s model, the emerging roles AI creates inside organisations, the allocation economy thesis, and why the binding constraint on AI-native work is not technical skill but management skill.

Key ideas

Allocation economy. We are shifting from a knowledge economy (people paid to do tasks) to an allocation economy (people paid to direct agents). Management skills — evaluating output, scoping roles, giving feedback, holding vision — were rare (8% of workers are managers); they will become universal.
Compounding engineering. Every unit of engineering work should make the next unit easier. Manifests as encoding feedback as prompts, building automations that accumulate, and investing in infrastructure that multiplies future throughput. Coined at Every by Kieran and Nityesh. See Compounding Engineering.
Head of AI operations. A new role: person who builds the prompts, automations, and knowledge bases that encode the principal’s taste into the organisation’s AI layer. At Every, this is Katie Parrott. She is not a developer; she is a process and communication expert who expresses the CEO’s judgment in machine-readable form.
AGI as profitable infinite agents. Dan’s operational definition: AGI is the point at which it is profitable to run agents indefinitely without human oversight — using the Winnicott leash metaphor from developmental psychology. A child who has fully internalised the parent’s values no longer needs the leash. AGI is when the agent has internalised enough judgment that the economics of oversight disappear.
CEO as adoption predictor. The number one predictor of org-wide AI adoption is whether the CEO uses ChatGPT (or equivalent) daily. CEOs who do can drive excitement and set realistic expectations; those who do not produce either no adoption or unrealistic expectations.

Every’s operating model

15 people; nobody writes code manually.
Products: Cora (AI email assistant), Spiral (AI writing), consulting arm, daily newsletter.
Cora built for ~$300K all-in including salaries, by 2 engineers and 15 Claude Code instances. Not technically possible three years ago.
Head of AI operations (Katie Parrott) builds automations that encode Dan’s taste: e.g., prompt templates from recorded feedback sessions with young writer Alex Duffy, who made a year’s progress in two months by treating each feedback session as a prompt to encode.
Model stack: o3 for personal/writing tasks; Claude Opus (first model to genuinely judge whether writing is good); Claude Code for all engineering; Gemini for app internals (cost); Codex for one-off features.

Compounding engineering

Dan’s account of the system at Every:

“For every unit of work we do, we try to make the next unit of work easier.”

Concrete mechanisms:

Every feedback session becomes a prompt in the system.
Automations are built once and reused; they do not need to be rebuilt.
The head of AI operations role exists specifically to convert tacit taste into explicit, reusable machine instructions.

This is distinct from agentic engineering (which is about correctness at speed) — compounding engineering is about the engineering of the engineering system itself. See Compounding Engineering and Agentic Engineering.

Head of AI operations

Katie Parrott’s role at Every illustrates the emerging pattern:

Responsibilities: building prompts, automations, and knowledge bases that encode the CEO’s taste.
Not a developer role; requires domain knowledge, communication skill, and judgment about what the CEO actually wants.
Acts as translator between human expertise and machine process.
Dan predicts this becomes a standard role inside companies as the allocation economy matures.

See Agentic Engineering for Boris Cherny’s observation that everyone on the Claude Code team codes, and AI Engineering for Chip Huyen’s framing of AI Engineering as a distinct discipline.

AGI definition

Dan’s operational definition, via the Winnicott developmental psychology leash metaphor:

“AGI is when it is profitable to run agents indefinitely without human supervision.”

The leash in child development represents the parent’s values temporarily externalised. When the child has fully internalised those values, the leash is no longer needed. The leash does not disappear because the child became more capable; it disappears because the economics of oversight change.

Applied to AI: AGI is not a capability threshold (passing a test) but an economic threshold — the point at which the cost of human oversight exceeds its benefit.

Allocation economy thesis

Based on Dan’s article written ~2.5 years before this episode. Core argument:

Current economy: knowledge workers are paid to do tasks. Managers are rare (≈8% of workers).
Future economy: AI agents do the tasks. Humans allocate attention — deciding what to work on, scoping roles, evaluating outputs, giving feedback, holding vision. These are management skills.
Management skills will democratise: they were expensive to acquire because managing humans was expensive. Managing AI agents is cheap.
Skills that become more valuable: evaluating output quality (taste), scoping and sequencing, vision (knowing what to want), knowing when to step in vs. delegate.

Generalists benefit disproportionately: AI provides on-demand specialised knowledge, so generalists who move across domains can stay generalist longer. Every deliberately hires and keeps generalists.

See also the Athens analogy: citizens of Athens were generalists (fighter, judge, juror, general) until empire forced specialisation. AI may reverse this.

AI adoption in organisations

From Every’s consulting arm (now ~$1M/year, growing):

CEO signal. CEO uses AI daily → org-wide adoption; CEO does not → near-zero adoption or unrealistic expectations.
Walleye hedge fund model (Dan’s reference case):
- AI-first memo written with ChatGPT (announced in the memo itself).
- Weekly meeting: people share prompts and use cases.
- Weekly stats email: usage numbers, celebrated early adopters.
The 10/80/10 rule. ~10% of employees are early adopters; ~10% will never touch it; ~80% will adopt if shown exactly what to do for their specific job. The consulting approach: customise training to each team, with exact prompts for exact situations.
Impact so far. “Further faster at the same budget” — not firing people, but increasing throughput without adding headcount.

Sip seed model

Dan’s funding innovation:

$2M committed from Reid Hoffman and Starting Line VC.
Drawn down on demand (the “sip”); not deposited in full.
Structured as a SAFE at a set cap.
Psychological benefit: can take risk knowing capital exists; does not face temptation to burn a large bank balance.
Philosophical alignment: Reid Hoffman “doesn’t care what size this business is” — aligned with Every’s mission (institution, not exit).

GPT wrappers as legitimate products

Dan explicitly defends the GPT wrapper as a valid product category, citing its malignment as wrong:

“GPT wrappers are amazing and they’ve been much maligned for absolutely no reason.”

The software-becomes-content analogy: ChatGPT is unbundling into specialised apps, just as Craigslist unbundled into many verticals. The wrapper is the product if it packages the capability for a specific context.

Writing as core identity

Dan stopped writing early at Every to focus on running the business; growth stalled and he was unhappy. Returned to writing as the centre of the company. The realisation: media businesses (newsletter + products) are structurally different from tech startups — the founder IS the product. References: Joel Spolsky (Trello, Stack Overflow), Jason Fried (Basecamp), Bill Simmons (The Ringer), Sam Harris.

Life motto: “Do things worth writing about, and write things worth reading.” — Pliny the Younger.