As coding agents become more capable, I’ve been wondering how the role of pre-built connectors may change.

Over the past few years, we’ve seen several engineering patterns emerge around agents: prompt engineering, workflows, skills, harnesses, and now loops. Each pattern has been useful in its moment, but the relative importance of these patterns seems to shift quickly as the underlying models improve.

In a recent agent benchmark, we observed that Skills still help in some cases, but the improvement was smaller than expected, and in some cases even negative. Agent Benchmark Blog

This made me wonder about a broader question:

How much should developer platforms invest in building connectors ahead of time, if agents are becoming increasingly capable of connecting services on themselves?

For example, SMS is a common requirement for many apps. InsForge could build a native SMS connector. But if a developer genuinely needed SMS, they probably wouldn’t wait for InsForge to ship one. They would have already opened a Twilio or Aliyun account, stored the API key in InsForge Secrets, and let an agent wire it into their app.

As agents get better at reading docs, they can literally integrate any external service they want, so what happens to the value of pre-built connectors? I used to believe its value because the glue code could save tokens and context window to make the agent more efficient. But this may no longer be true anymore, or at least the return is diminishing quickly.

Some questions I’m thinking about:

• When is a connector still meaningfully useful?
• When does it become something an agent can generate as needed?
• Are pre-built schemas, workflows, and integrations still important for saving tokens and context?
• Or does that matter less as agents get better at retrieval, tool use, and long-context reasoning?
• Which parts of a platform should be pre-built, and which parts should be left for agents to compose dynamically?

The question is like, if an agent could help build a car from raw materials, would people still buy a finished car?

The analogy is obviously imperfect, but it points to the same question: as agents become more capable, where is the boundary between useful pre-built infrastructure and work that agents can recreate on demand?

Let me know your opinions on this.