Agent-built SaaS replacements live or die on product polish

The pitch. Mahmoud Abdelwahab from Railway argues the "SaaS is dead" claim has a missing piece: most agent-built software works but does not feel good enough to use. He built an all-in-one internal marketing tool at Railway that replaces several SaaS products, managing website content, changelog, blog, email, social publishing, media, a mini CRM, task tracking, launch planning, calendar, comments, and realtime editing. He records the talk from a bench in Edinburgh and credits a "breakthrough" to GPT-6.

The prerequisite. His first principle is that the stack matters less than the codebase. He prefers TanStack for web work and Payload for content, but the real point is solid abstractions. An agent given a well structured repo will follow existing patterns instead of inventing new ones. That is plausible, but it assumes the repo is worth following, which is not true in many organizations.

The polish gap. The most useful content is a list of product patterns agents skip unless explicitly told. Autosave and protection against closing tabs, optimistic updates with rollback, undo after delete, archive instead of hard delete, URL state for filters and sorting, and infinite scroll instead of pagination. These are table stakes in commercial SaaS, and they make a good checklist for any team building internal tools with agents.

The skills layer. He packages design opinions as agent skills, covering UI, typography, layouts, error messages, animations, and Railway deployment. Telling the agent to use existing pages as references improves consistency. This is a practical way to encode product standards, though he admits the agent sometimes ignores the instruction. The realtime collaboration feature took a joking "700 million tokens later," a reminder that polished features still carry cost.

The verification loop. He calls verification the most important step. Agents say the build and tests pass, but the actual experience can be broken. He uses computer-use-based end-to-end testing so the agent can run the app, spot failures, fix them, and retry. This is the strongest operational idea in the talk, but it implies a serious compute and setup commitment.

The missing evidence. The talk is an engineering opinion, not a benchmark. There are no user counts, maintenance hours, downtime numbers, or cost comparisons against the SaaS products it replaces. The one-shot SaaS replacement is more credible than before under the right conditions, but "become your company's SaaS layer" is still a stretch for most organizations.