GitHub Issues and Codex agents now run in a single workflow
I now have an important practical result in agent-assisted software development.
For some time I have been working on bringing GitHub Issues and Codex agents into a single controlled workflow. The idea is simple: a task starts life as a normal issue in GitHub, and from there it moves forward not through manual handoffs between people and chats, but through a set of agent configurations, rules, and prompts.
Within that process, different agents handle the task step by step: intake, admission, planning, implementation, pull request creation, validation, merge decision, and issue finalization.
I connected this workflow to the repository viktor-gusev/dnd-game-master. Its issues are now flowing through that process. You can already inspect concrete examples directly on GitHub: #17, #19, #21.
For me, this matters not only as a technical experiment. Since February I have been working through the idea that users should be able to influence the code of their applications more directly, but not through a chaotic stream of prompts to a model. That influence should pass through a controlled process in which the product owner defines the approved boundaries of change in advance.
That idea has now taken shape as a working system you can actually inspect. It confirms my understanding of how people and language models can work together in software development: not as occasional participants in isolated chats, but as participants in a reproducible process with roles, constraints, and verifiable outcomes.
I described the broader model separately in the publication Turning User Intent Into Controlled Product Evolution. It serves as the conceptual backdrop for this post: what was still a general idea not long ago has now started working in practice.