Replacing Codex without replacing Kogen
Kogen currently uses Codex to write code during a Build. I want to try Pi and OpenCode too, but swapping the coding agent should not mean rebuilding Kogen around whichever one I choose.
That gets difficult if the coding agent also keeps the only record of the work, decides which checks count, reviews its own changes and tells Kogen when the job is finished. A different agent will expose different commands, events and ways of continuing a task. Kogen would have to learn a new software-building process for every integration.
I want Kogen to give the agent one approved change to implement, then keep the rest of the process under its own control. Kogen records the work, runs the final checks, sends a failure back to the same attempt, asks a separate agent to review the result and keeps the evidence afterward. The coding agent can still use all of its own tools while it works. It just doesn’t get to define when the Build is done.
I am going to test Pi and OpenCode on the same small repository and the same task. The first version will fail a check on purpose. Kogen should be able to reject that version, return the failure to the agent and continue the work without losing what happened before it.
That test will not tell me which coding agent is best. It will show whether I can replace Codex without also replacing Kogen’s checks, review and history. Neither agent has been selected, and I am not building an adapter yet.
← All notes