The software factory is the next extinct wrapper: within 3–6 months, better coding agents will absorb much of its core orchestration value. The future is in a slimmer skills based approach with guidelines, guardrails etc. https://x.com/i/article/2119166757625843712
The software factory is the next extinct wrapper
Factory AI, Droid and Warp are making a bet that the orchestration layer becomes more valuable as agents improve. I'm making the opposite bet: the model gets more capable, the wrapper gets thinner, and the skills and standards that matter most remain yours.
The value sold as a software factory will disappear through the next generation of coding agents within 3–6 months. These products may still be around, but as their core orchestration value fades, most people shouldn't need to use them.
Factory AI and its Droid product, and Warp Factories, are building systems for coordinating agents across the software development lifecycle. Factory describes an interconnected feedback loop. Warp describes a foreman that routes work through planning, implementation and review. My bet is that the agent increasingly handles that coordination itself.
Example of the current factory debate
Lauren Tan says you don't need a software factory. Eno Reyes, Factory's CTO, argues that a factory standardizes how work is validated. Their exchange captures the disagreement.
A factory can route work through planners, implementers and reviewers, moving context between them. My bet is that improving agents absorb more of that coordination. The company still sets its standards and decides what gets released.
Steve Yegge's bespoke factory
Steve Yegge created Gas Town and helped develop Gas City. He now says he has given up on reusable harnesses and is focused on Wyvern, his game, using a bespoke system called Wheelhouse. His essay “Fences, not Sandboxes” describes project-specific rules and their enforcement. His Wheelhouse account starts with Markdown files and coding agents, then builds a system around his game.
Wheelhouse is a substantial factory. What interests me is that it was built around a specific product and the decisions its creator wanted agents to follow. Those decisions are valuable because they belong to the work.
Yegge calls another part of Wheelhouse a Wish Factory. In Wyvern, admins and players can make requests that agents investigate, triage and turn into changes. He credits Guy Podjarny with the idea. That captures the feeling I'm seeing in my own work.
My wish factory granted WorldAI on iPhone
I've been building WorldAI, an AI RPG, with coding agents. My wish was to bring my existing desktop game to iPhone. The result is a whole iOS frontend, now available on the App Store. The iOS push took a few days of my hands-on time and about 1–2 weeks end to end.
I work part-time, roughly 20–30 hours a week. For about a week, my workflow was merging around 100 PRs a day across the work, about five times my previous pace of roughly 20 PRs a day. With the latest frontier models, it feels like I can make a wish and watch things happen. The experience is massively different from even three months ago.
In WorldAI for iPhone, a player can set up a campaign, see a character portrait and play through a scene. The app builds on my existing game and backend. The goal was to make the experience usable on an iPhone.
I use ChatGPT Dot alongside Claude, Codex and Grokbot, with my skill based PR workflow called /prw. I set the outcome and what must not break; agents implement, test and capture evidence. The work can start with a natural-language request, carry the relevant context to agents working in parallel, and come back with tests, review and evidence for me to inspect.
I review prompts, change size and proof, question choices and steer the work. As my Idea Velocity deck puts it: front-load alignment, back-load verification. I still make the product decisions and decide what merges for larger changes.
Deterministic checks are not deterministic thinking
Eno raises the risks of automerging in safety-critical software. But the choice isn't between buying a factory and giving up discipline. A company can encode its own standards in shared skills, guidelines and guardrails for whichever agents it uses.
Keep permissions and security boundaries enforceable. Run tests, check real user journeys, require evidence and keep merge and release authority clear. These checks don't require prescribing how an agent must reason through every problem.
The durable assets belong to the company
The knowledge and standards I expect to survive model upgrades are specific to the organization.
Infrastructure still has a job: keeping work running, isolating environments, recording results and connecting systems. I want company judgment in our own skills and tools, while agents take on more generic coordination.
The test is subtraction
After a model upgrade, try removing a layer on comparable tasks. Check whether the simpler setup meets the same requirements and evidence standard, what it costs and what failures it misses. Keep steps that improve outcomes; remove those that don't.
Sources:
PR workflow skills
Quick start: https://github.com/jleechanorg/jleechan-skills/tree/main#pr-workflow-entry-point
Main skill: https://github.com/jleechanorg/jleechan-skills/blob/main/.claude/skills/pr-lifecycle/SKILL.md
Idea Velocity deck: https://docs.google.com/presentation/d/1qLdMzZaIuP9IdpTBomuOGJTAspCEJo6Esu10zWFrhm4/htmlpresent
Factory's software-factory description: https://factory.com/news/software-factory
Warp's factory workflow: https://docs.warp.dev/factories/how-factories-work/
WorldAI Wish Machine examples: https://www.reddit.com/r/WorldAI_RPG/comments/1xnvucq/building_worldai_with_an_ai_workflow_that_now/