Shame to see it's all Nvidia tech. If you want it on AMDGPU as well I've got that part of the stack running. Can probably do something collaborative.
This is confusing as mostly the problem has been solved - suboptimally but "well enough" - by a patchwork of solutions every repo builds over time:
1. README.md - which agents can read and follow just fine. And they can debug the problems and if you ask, they can update the README so the next agent (or human) can follow the README with less trouble.
2. Dependency versions can be pinned in Dockerfile to avoid dependency hell.
3. Version pinning is supported by project.toml, Gemfile, package.json etc. as well. And all the major languages offer per-directory isolated dependency environments (uv/venv, nvm, rvm, chruby, etc.) So even non-containerized dev has no dependency hell.
4. Secrets for the "dev DB" etc. can be stored in a password vault or wiki. Which agents can access via a browser or curl, given your credentials - or else you can copy-paste them over.
Most crucially, development environments are only set up once when starting work on a new codebase. As rarely as once every few years.
Can you explain what's unsolved?