Dispatch from the Future: Building an AI-Native Company (Category AI)

I watched an ambitious, somewhat grandiose talk by Dan Shipper, co-founder and CEO of Every. He launched it with Nathan Baschez in 2020. Every now has 15 people running 6 business units and 4 products, publishing a daily AI newsletter and providing consulting.

The company is small, but its claims are ambitious and interesting. Here are the points Shipper makes:

1️⃣ A quantum leap at 100% adoption He argues that the difference between 90% and 100% AI adoption among engineers is not 10%, but 10x. If even 10% of the team uses traditional development methods, the organisation falls back on old processes. In this view, the mechanics of development change only with a complete transition.

2️⃣ One developer, one production application Each of Every’s four products was built by one engineer. These are not toys: Cora, an AI email assistant, handles thousands of inboxes; Monologue, a speech-to-text tool, is used to write millions of words a week; and Spiral, an AI writing partner, produces content with millions of impressions. Shipper says AI agents write 99% of the code, with nobody writing it by hand.

3️⃣ From the code editor to delegating to agents The key change is a terminal-based Claude Code workflow that replaces the traditional editor with task delegation. Developers can run 4 agent windows simultaneously, working on different features.

4️⃣ Demo culture versus memo culture When code is cheap, a company can move from writing documents to persuade colleagues to building and showing a prototype in a couple of hours. This makes room for stranger, more interesting ideas that are difficult to describe but easy to experience.

5️⃣ Compounding Engineering Every’s methodology aims to make each feature simplify the next one rather than complicate it. The cycle has 4 stages:

• Plan (40%): agents study the codebase and produce detailed plans.
• Work (20%): agents write code and tests.
• Review (20%): assess quality through tests, code review and subagents.
• Compound (20%): encode lessons in prompts, subagents and slash commands.

6️⃣ Secondary effects Shipper describes less obvious benefits of full adoption:

• Tacit code sharing: agents read neighbouring project repositories and transfer patterns across stacks without explicit shared libraries.
• Productive newcomers from day one: organisational knowledge is encoded in claude.md files.
• Cross-app commits: developers fix bugs in other products because doing so is easy.
• A polyglot stack: each product can use its own language and framework, with AI handling translation.
• Managers commit code: even the CEO can contribute production code between meetings.

7️⃣ Fractured Attention Programming Instead of needing a 3–4-hour focus block, a developer can leave a meeting, assign work to an agent, attend another meeting, return to the result and open a PR.

He describes the following implications:

1) Changes in development economics

• Parallel rather than sequential work: an engineer works on 3–4 tasks in separate worktrees instead of one at a time.
• Lower starting costs: prototype-first work becomes dominant.
• A changed developer role: agents write code and developers become orchestrators.

2) Process changes

• New building blocks: agents.md files with project context, specialised subagents and similar tools encode organisational knowledge.
• From documentation to artefacts: agents read code and these other building blocks directly.
• Different hiring needs: onboarding no longer takes weeks, and knowledge of one particular stack matters less.

#Engineering #AI #Metrics #Software #DevEx #Productivity #DevOps #Architecture #Culture #Engineering #ML #SystemDesign