@banteg
Post
omp has a very interesting eval/repl tool design. it runs a persistent ipython kernel pinned across session. this means imports, vars, open files survive even across subagents.
the agent can build the state incrementally like you would write a jupyter/marimo notebook yourself instead of rederiving everything every call or editing one-off scripts.
haven't seen this in any other harness. i think this design is the future and interleaving iterative work with thinking is a natural fit for agents.
Image from X post
Explanation
What it says @banteg argues that OMP’s standout agent design is its persistent IPython eval/REPL: one kernel stays alive across the session, so imports, variables, open files, and computed state survive tool calls and even subagents. That lets an agent accumulate working state incrementally, more like interactively building a Jupyter/Marimo notebook than repeatedly reconstructing context or writing disposable scripts. He thinks this pattern is likely to become standard for coding agents.
Context The screenshot shows the pattern in practice: successive Python cells download disk images, extract Fairytale filesystems, locate 503 `.MES` files across four games, then run a threaded detector and aggregate engine classifications. Each step builds directly on variables and filesystem state created by earlier steps.
Why it matters This attacks a real weakness in tool-using agents: repeated serialization, re-parsing, and re-derivation of intermediate state. A persistent interpreter can make exploratory coding, reverse engineering, data analysis, and debugging faster and more natural, while also giving subagents shared computational memory rather than just shared text context. The tradeoff is that hidden mutable state can make runs less reproducible and errors more path-dependent, so good checkpointing/state inspection becomes important.
Images The image materially demonstrates the claimed workflow: multi-stage analysis proceeds through persistent Python state rather than isolated one-shot scripts.