@VikParuchuri I need to create a benchmark for this at some point to incentivize ocr model makers to spend time on this :) (very important in legal domain!)
It’s still alpha, but already useful in local dev workflows.
If this sounds useful, I’d love people to try it and tell me what works and what doesn’t.
And @fishnets88, curious whether there’s a marimo-shaped version of this idea too?
This resonated enough that I went and built my version of it: agentnb.
It gives coding agents a persistent Python REPL they can drive through simple CLI commands, so they can iterate in-memory instead of constantly bouncing through scripts and restarts.
@lateinteraction
Though bash is a completely valid REPL, the amount of time coding agents lose during experimentation because they iterate on scripts instead of a Jupyter-like in-memory REPL is basically dumb.
Fixing 1 local bug should not require restarting the whole job. Need better scaffolds.
I think this is complementary to notebook-native approaches.
For example, @HamelHusain built a version of this idea where the agent works on actual notebooks that humans can inspect.
agentnb is more about giving the agent an in-memory Python workspace through small CLI calls.