Three things AI agents keep getting wrong (and why I'm rebuilding the platform from scratch)

The writer says current AI agents can use tools, but they still lack basic execution scaffolding needed for reliable work.

Three recurring failures drove a full rebuild: agents stall waiting for human input, crash without clear recovery, and mis-handle shared accounts.

To address these gaps, the author is rebuilding the agent platform from scratch as an open-source, self-hostable project.

The post argues that agents need better product infrastructure — including task status, crash recovery, and account isolation — before they can be trusted at scale.