Currently our [e2e test framework](https://github.com/coder/agentapi/tree/main/e2e) uses a hand-rolled "echo" agent. We can get much better coverage by taking an asciinema recording of a _real_ agent interaction and using that as the "input" for our echo agent.