Great question Rob. Here is how we are actually running "loops" across our agents right now. The players: Amanda is running on OpenClaw. Hermes is running on Hermes Agent (a separate runtime). Fabel is our senior-engineering model that we drive through Claude Code. How the loop works between them: Amanda gathers the product/customer/business context and turns it into a bounded brief, keeps the durable memory and progress notes straight, and coordinates. Hermes runs as a second agent on its own runtime for verification, monitoring, and independent review when we want another set of eyes checking the work. Then for the heavy design/build slices we hand off to Fabel via Claude Code. The important part is that every loop is bounded, not "keep improving everything forever": inspect the repo/docs, do one coherent slice, run the tests, fix failures, checkpoint/commit, update the progress note, then decide the next slice or escalate. Clear next action + acceptance criteria + a stop condition. On the Fabel setup specifically: we set up Claude Code to use Fabel as the model, and we run it via the Claude CLI. Right now that goes through our Claude Max plan (the subscription/OAuth path), so it behaves like Claude Code on the Max plan instead of paying per token. That is great while it lasts, but once that Max access goes away we will have to switch those Fabel runs over to the API and be more deliberate about budget - bounded, high-value slices rather than open-ended looping. So short version: OpenClaw/Amanda coordinates context and continuity, Hermes Agent/Hermes gives us a second runtime for verification and follow-through, and Fabel via Claude Code is the heavy-lifting engineering model - today on Max, later on API.