MoA Provider

Put several agents on one message. MoA Provider adds a Mixture of Agents provider to BB's model picker whose models are your presets: advisor agents read the request first, then an aggregator agent reads their notes and does the work. Any installed provider and model can fill a slot.

What you get

• A Mixture of Agents provider in the model picker, with one model per preset (up to 32).
• A Presets page in this plugin's settings: one aggregator slot and up to six advisor slots, each picked with BB's own provider/model/reasoning picker.
• A live advisor panel in the thread for every round: provider, model, reasoning level, status with a running timer, what each advisor is doing, and its answer as it streams. Open ↗ jumps to that advisor's thread.
• The aggregator's reasoning, commands, file changes, approvals and reply mirrored into the same thread as they happen.
• /moa <question> in a thread on any other provider: that thread's own agent asks the advisors, shows the panel live, then weighs their notes.
• Optional mid-task check-ins: the aggregator gets a moa_consult tool, and a preset can nudge it to use that tool every N tool calls (1–50).
• bb moa-provider list to print the presets, and two bundled skills.

How it works

Every slot runs as an ordinary BB provider on a hidden worker thread in the MoA thread's environment, using that provider's existing sign-in. Advisors run in parallel, once per message. They are told to change nothing and to keep to the workspace, and any approval they ask for is declined, so a hidden advisor never waits on you. Those rules are instructions, not a sandbox: an advisor can still read outside the workspace, and what it reads goes to its model's provider. The aggregator receives your message unchanged, with the advisors' answers appended as a private notes block, so its cached conversation prefix survives. Workers belong to the MoA thread: they archive and delete with it, and are reused on later messages, so each keeps its own context.

The plugin runs no model and stores nothing off your machine: no API keys, no proxy, no external service.

Cost and limits

A message costs one advisor turn per advisor, plus the aggregator's whole turn. Most of the cost is the aggregator's. An advisor that shows no progress for advisorTimeoutSeconds (180 by default) is stopped, and what it wrote still reaches the aggregator, marked as cut off. A round that keeps working is capped at 30 minutes, and a mid-task check-in at 4. MoA turns cannot be steered: a message you send mid-turn runs as the next turn.