Extensible Software in the age of LLMs

gist

Jeremy Morrell argues that LLMs let users author the long tail of software features that a static web product cannot ship. The pattern is a solid, accountable core plus capability-based sandboxes that run user or LLM-written extensions cheaply and safely. Webhooks set too high a bar, and handing API keys to user code leaks them; object capabilities do not. Salesforce proved multi-tenant programmable platforms at scale. In 2026, V8 isolates, microVMs, WASM, and especially Cloudflare Dynamic Workers make the same idea cheap enough for ordinary web apps.

ideas

• Serve the long tail by extension. A static product covers the head of demand; LLM-authored hooks can fill personal leftovers without bloating the core UI.
• Keep a solid, accountable core. Users get superpowers only if the host remains the trusted surface for auth, data, and compliance.
• Pass capabilities, not keys. Untrusted code should receive a narrow function, never a credential or raw fetch it can exfiltrate.
• Cheap isolation is the constraint. Thousands of snippets need idle-cost near zero, millisecond starts, hard limits, and tenant isolation.
• The web should not be left behind. Local IDEs and game mods already do this; Salesforce proved the hosted version, and Dynamic Workers make it newly cheap.

quotes