I’m super early stages with this project and exploring what it would look like agents were just unix programs running securely along side other Unix programs. A big benefit of that is I can pipe one agents output into another agents input.

Another idea I’m exploring is having agents write code for simple tasks. First example, if you run “hello world” with agent script, the first time an agent reads the text file, understands its simple, then writes a static memory.js file. When the program runs again, it hits that file and never invokes the agent.

Where things get interesting are the cases where a script works for some cases, but not all. When that happens, the memory.js script can call agent.resume() during a static run so the agent can take over.