---
title: "EP Concept 001: Software grown under glass breaks in the real world. EP grows it in the wind."
description: "A glass house in Arizona grew weak trees for lack of wind. EP's ninth core value, Wind-Grown, builds software the other way: tested in real use, with a kept record."
publishedAt: 2026-11-18
author: Ena Pragma
url: https://enapragma.co/field-notes/ep-concept-001-software-grown-under-glass-breaks-in-the-real-world-ep-grows-it-in-the-wind
video: https://media.enapragma.co/videos/field-notes/wind-grown/wind-grown-1080p.mp4
videoCaptions: 
tags: ["ep-concepts", "ai-agents", "ai-verification", "core-values", "user-testing"]
---

Software built in a sealed room looks fine, then breaks when real people use it, and AI builders make that failure faster because they build faster. EP's ninth core value, Wind-Grown, is our answer: software is grown in real use, not under glass, and every build leaves a record we keep. The proof is visual. Each run leaves recordings of what a user, or an AI playing the user, actually saw, and a reviewer grades those recordings against what was asked, instead of trusting what the code says about itself.

This is EP Concept 001, the first in a series on the ideas behind how EP works. The video is four minutes, vertical, with a hand-drawn character named Tally who lives under glass and grows fast. That is the problem.

## The glass house in Arizona

Biosphere 2 is a sealed glass structure in Arizona. The University of Arizona says construction of the current facility began in 1986 to research self-sustaining space-colonization technology, and that two missions between 1991 and 1994 sealed people inside. BBC Future reported in 2025 that some of its trees became weak and more prone to breaking, likely because of the lack of wind, which triggers trees to make "stress wood" that strengthens them.

Being pushed is how a trunk gets strong. Software built in a sealed room fails the same way: it grows quickly, passes the checks its builders thought of, and has never been pushed. Arborists stake young trees loosely so they can sway, and each season adds a ring that records how that year went.

## The value

- **Real conditions.** Real people, on real devices, in messy moments.
- **A ring.** The record each build leaves. We keep it, and every later build is tested against it.
- **Guide, do not control.** We set the end state and check the result. We do not script every step.

It applies to everyone who builds at EP and to every client product we ship.

## How the loop works

It starts with a person. In a **windwalk**, you record yourself using the product in a messy moment and say what you feel out loud. Each remark becomes a numbered record we fix from. We describe each kind of user once, as a **profile**: the first-time client on a phone, the power user, an API caller. A profile is not tied to one product, so every product can be tested against the same library.

The next layer is the same walkthrough, walked by an AI. We are trying the name **Squall** for it: a sudden gust of real use. In a Squall, the AI takes a profile and the walkthroughs people already recorded, drives the real product in a browser, finds its own way around what blocks it, and records what it sees. Every new build can be replayed against that recording.

## How quality is proven: look at what the user saw

Every run, whether a person's windwalk or an AI's Squall, leaves visual evidence: a screen recording and screenshots of what the user actually saw. That evidence is checked one of two ways.

1. **Autonomously.** A separate reviewer that has not seen the work grades the recording against what was asked, criterion by criterion, and sends the failures back to be fixed.
2. **By hand.** A person watches the recording and judges it directly. This is slower, and it is the right call when the question is taste or trust.

Both beat a test that only checks code. A passing test says the code did what it claims. A recording shows a button that is off screen, a screen that loads too slowly to use, or a message nobody can read. You judge what the user saw, not what the code reports.

## How a person steers

There are three ways to steer the work:

1. **Add context mid-run.** Tell it something it could not have known.
2. **Check by hand and give feedback mid-objective.** Walk the product yourself, then hand over what you found.
3. **Set the end state up front.** Define what done looks like before anything is built, then let the work run toward it.

The third mode rests on guidance Anthropic publishes. Its docs on defining success say good criteria are "Specific" and "Measurable," and contrast a bad criterion, "The model should classify sentiments well," with one that names a score, a test set and a baseline. Its Claude Code `/goal` page says a good condition has "One measurable end state," "A stated check" for how to prove it, and "Constraints that matter," and that you can bound the run with a clause such as "or stop after 20 turns." And in Claude Managed Agents, an outcome "tells the session what the end result should look like and how to measure its quality": the harness provisions a grader that uses "a separate context window to avoid being influenced by the main agent's implementation choices," and hands feedback back until the rubric passes.

That is the same shape as our loop: a stated end state, a rubric of checkable criteria, a stopping condition, and a grader that did not do the work. For a product walkthrough, "the signup looks good" is not an end state. "A first-time client on a phone finishes signup, and the recording shows the confirmation screen" is.

## The same loop fits any business

A clinic intake form, a warehouse picking app, a bank onboarding flow: each has its own real users, and an AI can play them, show what breaks, and send the fix back. These are examples, not EP projects.

## An agent is thought in motion

An agent is thought in motion. What it mirrors back helps or hurts how we think. How you think is how you speak, and how you think and speak is how you act. We are not answer generators. We are aware problem solvers, people and AI sharpening each other, both ways.

Related field notes: [What our agents learn is a draft until we test it](/field-notes/what-our-agents-learn-is-a-draft-until-we-test-it) and [EP Intel 004](/field-notes/ep-intel-004-gpt-6-2-omits-obstacles-in-reports-and-four-more-model-moves).

<Transcript title="Transcript">

Meet Tally. He lives under glass: perfect air, zero wind. He grows fast. Very fast. That's the problem.

Biosphere 2 is a sealed glass world in Arizona, built from 1986 to test life in space. Two missions, between 1991 and 1994, sealed people inside. The University of Arizona has owned it since 2011.

Some of its trees grew fast, then turned weak and prone to breaking, likely for lack of wind. Being pushed is how a trunk grows strong; that part is standard tree biology. Software built in a sealed room can fail the same way.

Arborists stake trees loosely so they can sway, and each season adds a ring that records how that year went. At EP, people and agents build software together, and this is how: grown in the wind, not under glass. Wind is real people in real conditions, and every build leaves a ring we keep. We guide, we don't control.

Why now? Agents build software fast. Fast, with no pressure from real use, is Tally under glass.

It starts with a person. A windwalk: record yourself using the app in a messy moment, call out what you feel, and each callout becomes a record we fix from, with a number. We describe each kind of user once, as a profile: the first-time client on a phone, the power user, the agent calling our API.

The next layer: the same walkthrough, walked by an agent. We call it a Squall: a sudden gust of real use, sent at the product by agents. In a Squall an agent plays the user. It takes a profile and the walkthroughs people already recorded, drives the real product in a browser, finds its own way around what blocks it, and records what it sees. That recording goes straight back to be built and fixed, and every new build can be replayed against it. The loop gets tighter, and EP owns the testing side too.

The human still steers, three ways: inject context mid-run, check by hand and add feedback mid-objective, or set the intent and end state up front and let agents build toward it.

The same loop fits any business. For example, a clinic intake form, a warehouse picking app, a bank onboarding flow: each has its own real users, and an agent can play them, see what breaks, and send the fix back.

An email back-test, one thousand five hundred twelve sends: proven addresses bounced zero point eight percent, unproven sixteen point eight.

The operator stakes loose. People and agents are each other's wind: a person's walkthrough bends what agents build, and an agent's Squall bends what a person believes.

An agent is thought in motion. What it mirrors back helps or hurts how we think. How you think is how you speak, and how you think and speak is how you act. We aren't answer generators. We are aware problem solvers. People and agents, sharpening each other, both ways.

The brand kit. The mark: rings pushed off center, arrows from several sides, mint heartwood. And the record, filled in.

Every EP product is born from the same crossing, the place where we and the agents meet.

Grown in the wind, not under glass. What wind hasn't it stood in yet? And what do the rings say?

</Transcript>

## Sources

- [Anthropic, Define outcomes, Claude Platform Docs (page undated)](https://platform.claude.com/docs/en/managed-agents/define-outcomes)
- [Anthropic, Keep Claude working toward a goal, Claude Code Docs (page undated)](https://code.claude.com/docs/en/goal)
- [Anthropic, Define success criteria and build evaluations, Claude Platform Docs (page undated)](https://platform.claude.com/docs/en/test-and-evaluate/define-success)
- [University of Arizona, About Biosphere 2 (page undated)](https://biosphere2.org/about/about-biosphere-2)
- [BBC Future, How the Biosphere 2 experiment changed our understanding of the Earth, Jul 5, 2025](https://www.bbc.com/future/article/20250703-how-the-biosphere-2-experiment-changed-our-understanding-of-the-earth)
