How agent harnesses actually work, what breaks when a run outlives its browser tab, how credentials should be scoped, and an honest account of what does not work yet.
Everyone building agents has access to the same models. What separates an agent that works from one that demos well is the harness around it: tool selection, memory, credentials, and a runtime that survives contact with reality.
KavelaReadAn honest list of Kavela's current limits: what is genuinely built, what is rough, and what we have not started. Written because every agent platform demo looks identical and none of them tell you this.
KavelaReadWhen you publish an agent other people use, whose account does it act on? There are exactly two right answers, and letting a tool sit in both states at once is how credentials leak between users.
KavelaReadMost agent frameworks tie a run's lifetime to the browser tab that started it. Here is why that breaks long jobs, and how Kavela runs agents in a durable object so closing the tab does not kill the work.
KavelaRead