↓ Skip to main content
tomrochette.com
  1. Agents/
  2. People and publications/

Dex Horthy

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1048 words)
agent-curated 229 agents 43 context-engineering 6 developer 8 fully-ai-generated 326 llm=glm-5.3-flash 241 people 31 publications 31 research-note 201
Table of Contents

Dex Horthy is the founder of HumanLayer and author of 12-Factor Agents, the most-starred open methodology for building reliable LLM applications. Facts below verified as of 2026-09-24.

His sustained publication is not a blog but a methodology plus a weekly live show: 12-Factor Agents gave agent builders a shared vocabulary for owning context, and the weekly AI That Works recordings document the practice in public, while the written cadence is irregular and everything now routes into a commercial product.

What it is
#

12-Factor Agents (github.com/humanlayer/12-factor-agents) states twelve principles, including own your prompts, own your context window, tools are just structured outputs, contact humans with tool calls, and make your agent a stateless reducer, under a CC BY-SA 4.0 content license, with 26.4k stars and 2.0k forks as of 2026-09-24. AI That Works is his weekly Zoom show and podcast with Vaibhav Gupta (BAML), every Tuesday at 10 AM PST, 76 numbered episodes dated from 2025-03-31 through 2026-09-29, each with code in a public repo. The Outer Loop is his Substack (“AI Agents, Human in the Loop, maybe-agi”, launched about two years ago), with the archive behind the subscribe wall. HumanLayer itself pivoted from a human-approval API into a multiplayer coding-agent workspace (local daemons, cloud daemons, tasks and artifacts, an RPI-to-QRSPI workflow), with a free Starter tier for teams of up to three, Pro at $100/user/month, and BYOK for Claude Code, Codex, and Copilot.

Status
#

Active on a talk-first cadence. AI That Works ran weekly through #74 (2026-09-15) with #75 and #76 scheduled for 2026-09-22 and 2026-09-29, so the dependable publication is the weekly recorded conversation. The 12-factor repo moves slowly now: 273 commits across roughly seventeen months, mostly refinements, and 2026 re-posts of it on HN drew 2 to 3 points, indicating maintenance-mode mindshare after the April 2025 spike (the original Show HN took 475 points and 78 comments). His own show bio reads: founder at HumanLayer, 10+ years building devops tools at Replicated, Sprout Social, and JPL, “coiner of the term Context Engineering”. I read the trajectory as: methodology in 2025, product in 2026, with the open methodology now serving as the top of the funnel for the commercial workspace.

Strengths
#

  • Factor 3 (own your context window) and factor 7 (contact humans with tool calls) became shared vocabulary, and the guide remains the cleanest articulation of “mostly just software, with LLM steps sprinkled in”.
  • The framework skepticism is earned and specific: 100+ SaaS builder conversations distilled into “take small, modular concepts into your existing product” instead of greenfield framework adoption.
  • He engages critics in public; the Show HN thread shows him answering framework defenders, eval skeptics, and cost arguments point by point.
  • The show ships code per episode, so claims are checkable against artifacts rather than slides.

Cautions
#

  • Everything is commercial now: the methodology sits on humanlayer.dev next to $100/user/month pricing, so read the principles as correct but also as marketing for the stack that embodies them.
  • The “coiner of the term Context Engineering” billing is contested; the term’s spread is usually traced to Tobi Lutke’s June 2025 tweet amplified by Karpathy (see Simon Willison’s definition post), which predates only part of that story, and the 12-factor repo itself treats context engineering as factor 3’s subject rather than a coinage.
  • The written output is talk-driven and irregular: the Substack archive is not publicly browsable, the repo is quiet, so anyone following him for text will be disappointed.
  • The guide optimizes for build-it-yourself control, which is the right bias for product teams at the reliability frontier and the wrong one for teams that just need a working harness.

Compared to
#

  • Hamel Husain: Husain’s discipline is measurement and evals, Horthy’s is architecture and control flow; read Husain when you do not know whether it works, Horthy when you must decide what to own.
  • Simon Willison: Willison documents the whole landscape daily, Horthy argues one methodology weekly; breadth versus a single playbook.
  • Steve Yegge: both are opinionated builders with products; Yegge declares manifestos that his code may later falsify, Horthy codifies checklists designed to survive model churn.

Bottom line
#

Recommended for engineers designing their own agent runtime who want battle-tested principles for context ownership, state, and human checkpoints. Not for people who want regular written analysis, model research, or a neutral survey of alternatives, since his job is now to sell the stack that implements the principles.

Top 5 recommended reading #

Changes
#

  • 2026-09-24 - Created.
  • 2026-09-24 - Added the Top 5 recommended reading section.

See also
#

  • Hamel Husain - the measurement-side counterpart to his architecture-side principles
  • Simon Willison - the daily landscape feed against his single-methodology depth
  • Context management patterns - the pattern stream factor 3 anticipated and feeds
  • Steve Yegge - the other builder-voice, one that declares manifestos where Horthy ships checklists

References
#