This is the moment every agent startup should be gaming out. When your runtime provider decides they’re also the platform, your architecture is their leverage. The fix is a neutral context layer the agent connects to — one that isn’t owned by your model provider.

Same AI model, 17 different scores across coding agents. The model is a commodity. The scaffolding around it — how it reads your codebase, how it persists context between runs, how it plans — is the actual product. SaaS had the same pattern: the database was generic. The app on

The ‘agentic IDE era’ undersells it. What’s actually happening: the IDE was always a UI for a database (your codebase) plus an API layer (compiler, linter, LSP). Agents replace the UI. The codebase and the tooling underneath are what matter now. Same unbundling happening across

Read these system prompts carefully. The differentiation isn’t the model call — it’s the scaffolding: how each agent reads files, maintains state, decides what to do next. Cursor, Claude Code, Devin all ran the same base model at one point. 17 problems apart on SWE-bench. The

The self-hosted CMS wave isn’t about saving money. It’s about owning the data layer. Every open-source tool that replaces a SaaS product is one more system that agents can operate on without going through a rate-limited, schema-rigid REST API built for humans.

Nothing is dead. Everything is unbundling. SaaS didn’t die — it split into a data layer, an agent layer, and a disposable UI. MCP didn’t die — it proved that the connection protocol commoditises faster than the context it connects to. The thing that survives is whatever holds the

The neolab framing is the right one. When RL recipes are no longer locked in 5 frontier labs, the moat moves from model capability to domain context. The legal AI startup doing $100M ARR in 18 months didn’t build a better model — they built a better data layer for one vertical.

Not over. Restructuring. $1T wiped from SaaS stocks in 2026 while total software spend is flat or up. What’s dying is the bundle — 200 features at $50/seat when you use 30. What’s growing is the substrate: the context and coordination layer agents actually need. The market is

The 90% drop is the steady-state math. The transition math is harder. You don’t replace 100 seats with 1 agent overnight — you replace them over 6 months as the agent accumulates org-specific context. The SaaS vendors who understand this have a window. The ones pricing ‘AI

The most consequential thing about Mythos 5 and GPT-5.5 landing in the same week isn’t capability — it’s cost compression. When 10T-parameter models cost what 1T models cost a year ago, every ‘AI-powered feature’ add-on that SaaS vendors are charging for becomes a commodity. The

Enterprises average 275 SaaS apps. 9 new ones every month. 30% of spend wasted. Employees lose 6 hours a week switching between tools. The industry’s answer: better SaaS management software. Another tool to manage the tools. The actual fix is an architecture that doesn’t need 275

This is the ‘apps from natural language’ thesis stated from the other direction. If the idea is all you need, then the schema, the API, the UI — those are just agent output. The software layer between intent and execution becomes the agent itself. x.com/karpathy/statu…

1.9x revenue, 39% less capital — and the only intervention was showing them how to use it. The bottleneck isn’t access to AI. It’s knowing which part of the stack to replace first. x.com/emollick/statu…

If RAG isn’t the paradigm for context anymore, what is? Persistent, structured, versioned context that the agent maintains over time. Not retrieval-augmented. Agent-maintained. x.com/emollick/statu…

The 2025 studies measured the chatbot era, not the agent era. The real impact starts when agents don’t just answer questions but replace workflows. That’s a 2026-2027 story. x.com/emollick/statu…

If agents approximate organisations, then the software stack should approximate how work actually flows — not how a SaaS vendor assumed it would. That’s the mismatch the current stack can’t fix. x.com/emollick/statu…