From a single pod to per-user runtimes
Why a multi-tenant worker pool was the wrong abstraction the moment users brought their own MCP servers.
Why a multi-tenant worker pool was the wrong abstraction the moment users brought their own MCP servers.
RLS, auth, and a Postgres we don't have to babysit. The trade-offs are real; the leverage is bigger.
Partial JSON is a UX problem first, a parser problem second. How we render incomplete tool calls cleanly.
Most retry libraries assume failures are transient. Real APIs disagree. A small taxonomy goes a long way.
The marketplace was the second-most-asked-about feature. Removing it made the product easier to explain.
Internal codenames leak into UI more often than you'd think. We audited ours and renamed half of them.
Settings pages start clean and end as a graveyard. We rewrote ours from scratch and gave it rules.
You don't need a thousand-row eval set to catch the regression that matters. You need the right twenty rows.
Most teams reach for fine-tuning too early and prompt engineering too late. Here's the heuristic we use.