Let's see, this is a great project, but the sandboxed virtual machine approach makes sense. I've had Claude Code destroy a configuration file on my local machine more than once, so running it remotely is actually smart, but I'm not sure.
However, when creating MCP tools, I've noticed that the processing isn't really what's expensive. It's when the agent gets stuck in a loop. For example, I note writes faulty code, executes it, detects the error, tries to "fix" it, makes it worse, rinses it, and repeats it 5 to 10 times. This quickly consumes the context window, regardless of where the code is running.
I'm curious: Does Oblien detect when an agent is simply wasting time? Or is that left to the agent framework? Because, I mean, sometimes I look for that solution and there isn't one :(
However, when creating MCP tools, I've noticed that the processing isn't really what's expensive. It's when the agent gets stuck in a loop. For example, I note writes faulty code, executes it, detects the error, tries to "fix" it, makes it worse, rinses it, and repeats it 5 to 10 times. This quickly consumes the context window, regardless of where the code is running.
I'm curious: Does Oblien detect when an agent is simply wasting time? Or is that left to the agent framework? Because, I mean, sometimes I look for that solution and there isn't one :(