This is how Scale By Video actually runs, the infrastructure, the specialist agents, and the weekly learning loop that turns every client's results into a sharper next round. Internal reference, current as of June 2026.
Each layer does one job. Paperclip decides and pays, OpenClaw runs the agents and touches the outside world, Hermes remembers and automates.
The orchestration and budget layer. Routes incoming work to the right specialist agent, enforces a per-agent spend cap, and exposes the budget API that Ledger reads. Every agent is invoked through it.
The gateway that runs the 14 agents as subprocesses and connects them to the outside world: channels (Telegram capture via Scribe) and MCP connectors (the official Meta Ads MCP, HubSpot, Slack, Calendar). Where the agents actually do things.
The NousResearch runtime. Holds Honcho-backed memory, runs scheduled jobs and skills, and provides native gateway prompt caching. The agents' long-running thinking and recall live here.
Cheap, deterministic work runs on Haiku; reasoning and client-facing work on Sonnet; only Boss's weekly synthesis touches Opus. That tiering plus prompt caching is what keeps the whole stack running on a few dollars a day.
Without it, every client engagement is a clean slate, the same patterns re-discovered and forgotten. The mesh is a closed weekly feedback loop on ad performance. This is the compounding edge, not the tools.
Every run starts with a vault + Honcho read and ends with a line in the daily note. Nothing is done from a blank slate.
The memory daemon crystallizes the day's transcripts and outcomes back into vault notes, so tomorrow's agents know more.
Signals become decisions become the next batch. Anchor closes the loop: did the decision actually improve booked → showed → closed?
This isn't a one-time setup that hands you a reel. It runs on three clocks, and the longer it runs the sharper it gets.
Scale By Video · internal field guide · sbv-prod-01 · brain synced Mac → VPS · current June 2026