INVALID SUBMISSION - cudex

Please ignore submission as I can’t find the conversation ID as I used Codex via T3 Chat. I’d delete my submission but you can’t for some reason.

Inspiration

Codex is powerful, but running agent workloads locally can consume resources and expose more of a developer’s environment than necessary. Particular risks we want to reduce are vulnerable dependencies, overeager agents, and data exfiltration.

What it does

This project forks Codex to ensure it always runs inside a sandbox VM hosted by CubeSandbox. Any subagents are instead running full agents in their own, fully isolated VMs. Code changes are fed back via git patches.

How we built it

This project relies heavily on open source solutions as well as Codex with GPT-5.6. I started with going through the idea with Codex to find out how feasible the idea is by cloning the Codex repo and going through the code. I then further discussed with Codex until I had a clear picture of what we’d need to do. I made have use of plans and a large amount of threads.

Challenges we ran into

The biggest challenge was and is the sandbox ownership model. I also decided to replace the optimized subagents design in Codex with ‘full’ agents to simplify the architecture for this project.

Accomplishments that we're proud of

Honestly just getting something working that would’ve been practically impossible just a year ago and still challenging with 5.5.

What we learned

A lot about the projects we rely on as well as agentic development with this level of model. “Just tell it what you want done, less how you want it done” is probably my most important lesson.

What's next for cudex

• simplified deployment, preferably with a skill-based setup
• improved logging
• hardened security
• custom agent templates including configurable file access, sandbox resources, and network filters

Built With

• codex
• cubesandbox
• e2b-sdk
• postgresql
• rust
• typescript