Platform comparisons in the agent space lean hard on connector counts — 400 integrations, 1,000 integrations, an MCP catalog. The number is easy to market and easy to audit, which is exactly why it’s worth asking what it actually measures.
A connector library is a moat only against users who can’t generate their own tools. With a working code factory, wiring any external API is a single prompt. Our data point: a complete new LLM-provider backend — auth, streaming, model listing, cost accounting — took one prompt and about fifteen minutes, and arrived as versioned, addressable, first-class code with its own endpoint and its own UI. Not an adapter conforming to someone else’s interface: a native artifact we own and can extend.
That last part is the underrated half. A generated integration is strictly better than a catalog connector along every axis except the first thirty seconds: it does exactly what your use case needs, it’s debuggable in your own codebase, and when the upstream API changes, the same factory that wrote it updates it.
So the axis that actually matters is time-to-first-working-integration — and the honest way to think about any platform gap is a two-bucket distinction:
- Missing-and-expensive: sandboxing, high availability, multi-tenancy. Boring, load-bearing infrastructure that no prompt will generate. These are real gaps, and we name ours plainly.
- Missing-and-one-prompt-away: most integrations, most format converters, most glue. Not gaps — latent capabilities with a fifteen-minute realization cost.
Only the first category should count against a platform. Judging a system with a code factory by its connector count is like judging a machine shop by its inventory of finished parts.
The residuals are real and worth naming: cold-start discovery means you must know to ask for an integration (nothing suggests one exists), and native tools are invisible to external MCP clients. We accept both — but they’re the actual trade, and catalog size isn’t.