Factory AI and Warp are betting that orchestration becomes more valuable as coding 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.
I expect much of the orchestration value sold as a software factory to be absorbed by 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 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.
My wish factory built an iPhone app
I've been building an AI RPG with coding agents. My wish was to bring my existing desktop game to iPhone. The result is a whole iOS frontend. 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. The app builds on my existing game and backend. The experience is massively different from even three months ago.
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 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. I still make the product decisions and decide what merges for larger changes.
Deterministic checks are not deterministic thinking
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.
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.
What parts of your agent workflow are durable engineering knowledge, and what parts could the next model upgrade make unnecessary?
Read the full article on X.
This article was prepared with AI assistance.
Top comments (0)