points by crisopolis 2 days ago | 0 comments
I've had the Claude Ultra tier for awhile and sometimes dabble in Codex. I also use OpenRouter for model access within the project not to build it.
When it comes to harnesses/orchestration I flip flop between raw Claude Code and subagents to my new favorite, OpenRig.
Where I have two rigs (Teams) which are Product and Engineering. Each of these rigs have seats for a model harness such as Codex or Claude Code. My product team is a Product Lead, Product Spec Reviewer. Engineering team is Lead, Dev, QA, Review. They send messages between themselves resolve Gitea Issues, open PRs.
And more or less yes, it's a software factory (I know I'm not gonna need it) but it's fun. I only talk to leads, mostly the Product Lead.
Use case I'm trying to build a SaaS with an Agentic core within the Staffing Industry.
Something I found really helpful for saas was a dogfood rig. It tests it as different kinds of user personas. Having a mix of blank slate agents (new user) and persistent agents (returning user) is nice. The persistent agents remember why things are as they are without having to look everything up and their feedback is generally much better than a blank slate user. Vercel's agent-browser has a good skill for this, it will find stuff I would have never thought to test myself.
That feedback is sent to a pm-intake seat, who checks it against the roadmap, etc (missions/slices). The product-lead will then talk to you about it. If they are creating a lot of feedback, something like a daily digest will keep it manageable so you can approve/deny in bulk. It might require a bit of tuning at first but at some point your confidence might be high enough that you just auto-approve entire classes of findings and only flag things are major product changes.