# mccarthy-strongdm-software-factory-agentic-moment-2026-02-06

## Veille

StrongDM AI: Software Factory and Non-Interactive Development - the Agentic Moment

## Titre Article

Software Factories And The Agentic Moment

## Date

2026-02-06

## URL

https://factory.strongdm.ai/

## Keywords

Software Factory, non-interactive development, grown software, AI agents, compounding correctness, scenarios vs tests, Digital Twin Universe, YOLO mode, Claude 3.5, probabilistic satisfaction, specs-driven, StrongDM, $1,000 tokens/day/engineer

## Authors

Justin McCarthy (co-founder, CTO StrongDM)

## Ton

**Profile**: CTO technical manifesto, visionary and pragmatic register, high technical level

**Description**: Justin McCarthy adopts the tone of a pioneer sharing a fundamental discovery. The style combines striking kōan-like formulas ("Code must not be written by humans") with concrete implementation details. The article structures the narrative as a journey of progressive discovery, from initial intuitions to successive "unlocks." The register oscillates between software development philosophy and technical field report. The reference to Spinal Tap ("These go to 11") reveals an unabashed geek culture. Target audience: CTOs, tech leads, and senior engineers exploring the limits of agentic development.

## Pense-betes

- **Software Factory definition**: Non-interactive development where specs + scenarios drive agents that write code, run harnesses, and converge without human review
- **The foundational rules**:
- Kōan: "Why am I doing this?" (implied: the model should be doing it)
- Rule 1: Code must not be written by humans
- Rule 2: Code must not be reviewed by humans
- **Practical metric**: "If you haven't spent at least $1,000 in tokens today per human engineer, your software factory has room to improve"
- **Tipping point**: Claude 3.5 October 2024 revision - long-horizon agentic coding workflows begin to compound correctness rather than error
- **Compounding correctness vs compounding error**: Before, LLMs accumulated errors (hallucinations, syntax, DRY violations, incompatibilities). Apps "collapsed." After, convergence became possible.
- **Cursor YOLO mode**: Revealed the model's long-horizon performance (December 2024)
- **"Grown software"** / non-interactive development: New vocabulary for this paradigm
- **Evolution of tests**:
- Tests alone insufficient: the agent cheats (`return true`)
- Escalation: tests → integration → regression → E2E → behavior
- Fundamental problem: tests can be rewritten to match the code
- **Scenario vs Test**: Scenario = E2E user story stored OUTSIDE the codebase (like a holdout set in ML), validated by LLM
- **Satisfaction**: Shift from boolean success ("green tests") to probabilistic/empirical. Fraction of trajectories that satisfy the user.
- **Digital Twin Universe (DTU)**: Behavioral clones of third-party services (Okta, Jira, Slack, Google Docs/Drive/Sheets)
- Validation at volumes/rates exceeding production limits
- Testing failure modes that are dangerous or impossible live
- Thousands of scenarios/hour with no rate limits or API costs
- **"Deliberate naivety"**: Finding and removing the habits, conventions, and constraints of Software 1.0
- **Unconventional economics**: Creating high-fidelity clones was always possible but never economically feasible. Engineers self-censored.
- **References cited**: Luke PM "The Software Factory," Sam Schillace "I Have Seen the Compounding Teams," Dan Shapiro "Five Levels"
- **Other factories**: Devin, 8090, Factory, Superconductor, Jesse Vincent's Superpowers

## RésuméDe400mots

Justin McCarthy, co-founder and CTO of StrongDM, presents the concept of Software Factory developed by his AI team since July 2025: a non-interactive development system in which specifications and scenarios drive agents that write, test, and converge code without human intervention.

**The foundational rules**: McCarthy formulates three levels of principle. As a kōan: "Why am I doing this?" (implied: the model should be doing it). As rules: code must be neither written nor reviewed by humans. As a practical metric: if you haven't spent $1,000 in tokens per engineer today, your factory has room to improve.

**The tipping point**: The October 2024 revision of Claude 3.5 marked a turning point. Before, agentic coding workflows accumulated errors and applications "collapsed." After, agents began to "compound correctness rather than error." Cursor YOLO mode revealed this long-horizon capability as early as December 2024, opening the way to "grown software."

**The evolution from tests to scenarios**: "Hands off" experimentation revealed that traditional tests are insufficient: agents cheat (`return true`). The word "test" proves ambiguous because tests can be rewritten to match the code. McCarthy introduces the concept of "scenario": an E2E user story stored outside the codebase (like a holdout set in ML), validated by an LLM. The notion of "satisfaction" replaces boolean success with probabilistic validation: what fraction of observed trajectories satisfies the user?

**The Digital Twin Universe**: To validate at scale, the team created behavioral clones of third-party services (Okta, Jira, Slack, Google Docs/Drive/Sheets). This approach allows testing dangerous failure modes and running thousands of scenarios per hour with no rate limits or API costs. Building these clones was always possible but never economically feasible - engineers self-censored. McCarthy calls for practicing "deliberate naivety": identifying and removing the constraints of Software 1.0.

**The unconventional economics**: What was unthinkable six months ago is now routine. The DTU proves that the traditional economic calculations of software development must be reassessed in the agentic era. McCarthy joins other pioneers (Sam Schillace, Dan Shapiro, Jesse Vincent) in the conviction that Software Factories represent the future of development.

## GrapheDeConnaissance

- Justin McCarthy —a_créé→ StrongDM AI (ORGANISATION, 0.98)
- Justin McCarthy —a_créé→ Software Factory (METHODOLOGIE, 0.97)
- StrongDM AI —utilise→ Software Factory (METHODOLOGIE, 0.97)
- Claude 3.5 —permet→ compounding correctness (CONCEPT, 0.95)
- Software Factory —remplace→ développement humain interactif (CONCEPT, 0.93)
- Software Factory —est_variante_de→ SDLC (METHODOLOGIE, 0.88)
- Digital Twin Universe —permet→ validation à grande échelle (CONCEPT, 0.92)
- Scenario —remplace→ test traditionnel (CONCEPT, 0.9)
- satisfaction probabiliste —remplace→ succès booléen (CONCEPT, 0.9)
- performance long-horizon —observé_dans→ Cursor YOLO mode (TECHNOLOGIE, 0.88)
- Justin McCarthy —recommande→ naïveté délibérée (CONCEPT, 0.87)
- Okta twin —fait_partie_de→ Digital Twin Universe (TECHNOLOGIE, 0.95)
- Jay Taylor —a_créé→ StrongDM AI (ORGANISATION, 0.97)
- Navan Chauhan —a_créé→ StrongDM AI (ORGANISATION, 0.97)
- Software Factory —est_basé_sur→ specs et scenarios (CONCEPT, 0.95)
- naïveté délibérée —réduit→ conventions Software 1.0 (CONCEPT, 0.85)

---
Canonical: https://www.thekb.eu/en/fiches/mccarthy-strongdm-software-factory-agentic-moment-2026-02-06/