I've been experimenting with systems built around AI agents recently and ran into a recurring problem: connecting agents to real-world services is still surprisingly fragmented.

Every integration tends to require its own authentication flow, token management, permissions handling, and API logic. After a few integrations the architecture becomes more about maintaining service connections than about the agents themselves.

Because of that I started building a project called Kaeso.

The idea is to explore a unified infrastructure layer where external services can be connected once and then accessed by agents through a structured interface. The goal is to simplify the integration side so developers can focus more on agent logic rather than constantly rebuilding connection layers.

While building it, the original idea actually changed quite a bit, which I wrote about here: https://kaeso.ai/blog/redefining-kaeso

The project is still early, but I'm interested in hearing how others here approach this problem.

If you're building systems around agents or automation, do you usually implement integrations separately for each project, or do you maintain some kind of shared integration layer?