Orchestration
One agent answers one prompt. Orchestration is what you reach for when one is not enough - when a problem wants several perspectives, a check on the first answer, or a pipeline of steps.
The AI Mastermind
Section titled “The AI Mastermind”The Mastermind runs several models against the same problem in one room. Beyond the plain modes, a mixture-of-agents mode has each participant answer, then converges the answers into a single result - and it estimates the cost before it spends it.
Per-agent model selection and a thinking toggle are configured per room, so you decide which family answers for each seat. The roster is filtered by the Model Manager, so an engine switched off there never takes a seat.
Failover vs. orchestration
Section titled “Failover vs. orchestration”These are different tools. Failover (in the control plane) keeps a single request answered when a provider is down - it is invisible and automatic. Orchestration deliberately runs several providers because you want more than one answer. Do not confuse the retry chain with the Mastermind.
Background work
Section titled “Background work”Longer runs - a translated document, a generated course, a rendered video - execute as background jobs. You can leave for another app and come back; the job keeps going, and there is a kill switch for runs you no longer want.
The orchestration surfaces
Section titled “The orchestration surfaces”M.A.V.E.N has several, each for a different shape of work:
- AI Mastermind - many models on one question (ASK / DISCUSS / MoA).
- Hermes Swarm - a visual graph of agents wired into a pipeline.
- Pipeline - an inbox-to-shipped flow.
- Kanban - the Hermes task board.
- Paperclip - an agent company.
- Loop - build → judge → fix, round after round.
In M.A.V.E.N Teams Loop, Pipeline and Kanban are one shape - Workflows - folded out of the shared event log.
- Local-first
- Guides - a real end-to-end example.