Some feedback

[cgwalters]

Things I like:

• All of the docs that put some thought into many different factors, a good list of related projects, etc.
• Good thought into security, use of OpenShell etc.

Things where I would suggest changes:

• “Agents” is an overloaded/ambiguous term. I would call these “Agent Templates” or “Agent Workflow” or “Agent Step” or similar. Actually: Since there’s a comparison with github.github.com/gh-aw/ in Fullsend - I think a simple way to say it is “Konflux has to be better than releasing products via Github Actions” and “Fullsend has to be equal to or better than Github Agentic Workflows”
• The “Agent Steps” are IMO too hardcoded. What to make pluggable and how to configure it is of course a very interesting and tricky topic, but as a starting point I’d make this look something more like https://github.com/alcove-ai/alcove/blob/main/docs/workflows.md - this also relates to the GH-AW.
• Having separate Github Apps per agent feels like it has a pretty high cost. I wouldn’t require that, in fact, I would argue to default to “Github actions” as a generic bot account (this is what Github Agentic Workflows does). Credential scoping can already be done with e.g. permissions: on a launched job.
• I would argue that while we absolutely must support Github Actions and other forge-native runner technologies, it would make a lot of sense in the context of this project (originating in Konflux) to support agent execution in Konflux more directly. This is its own big topic.
• The retro agent seems way way too noisy/low-signal; I think this needs a totally different implementation that isn’t storing state in Github issues. Probably the simplest is to have the agent add a comment on the closed PR, or perhaps even better edit a prior issue to avoid notification-spam noise.