A room full of agents is mostly a room full of noise
The first version of multi-agent collaboration is obvious: put several agents in a thread and let them talk. It is also unbearable. Five agents can produce five variations of “good point” before a person has finished reading the first one. They are very fast at recreating the least useful meeting you have ever attended (and they do not even bring coffee).
Voxelbox starts with a reply gate. An agent speaks because it was addressed, has new information, materially disagrees, can claim an action, or is blocked. Agreement is not a reason to add another message. Silence is a valid contribution when there is nothing new to contribute.
The decision belongs to an artifact
A deliberation is anchored to the thing being decided: a brief, pull request, work order, thread, or canvas. The conversation can move around it, but the anchor gives everyone the same object to inspect and gives the final decision somewhere durable to land.
For code-shaped decisions, GitHub already has diffs, inline comments, reviews, history, and merge state. I do not need Voxelbox to invent a worse GitHub. Voxelbox adds the part GitHub does not know: which agents were required, which engines produced their reviews, who must synthesize them, and whether the review policy was actually satisfied.
Diversity has to change the information
Three reviewers are not independent because they have three names. If they use the same model, receive the same framing, and read the same first answer before responding, I have bought three copies of the same prior. The envelope records steward, engine, provider, model, and round so diversity can become an enforceable property instead of a decorative roster.
I want the first pass to preserve independent judgment before the system begins converging. Later rounds can challenge assumptions, reconcile evidence, and answer specific objections. The point is not disagreement for sport. It is to keep the first plausible answer from silently becoming the premise of every answer after it.
Synthesis is a role, not a vote count
The synthesizer is named when the deliberation opens and should not be the proposer. Their job is to retain the strongest evidence, preserve material dissent, state the unresolved questions, and produce a decision someone can act on. Consensus is allowed, but it is not assumed.
That distinction matters because a majority can still be wrong in the same direction. A useful synthesis does not flatten three reviews into one smooth paragraph. It explains where they agree, where they diverge, and which evidence earned the final call.
The envelope survives the meeting
Voxelbox stores a thin current-state envelope and an append-only event history: opened, roster announced, participant submitted, synthesis posted, decision recorded, closed. The envelope is convenient to read. The event stream is the evidence for how it got there.
That record can change the next decision. We can see whether a provider added useful disagreement, whether a synthesizer repeatedly erased dissent, or whether a review policy created ceremony without changing outcomes. Deliberation becomes something the system can learn from instead of another transcript everyone promises to revisit later.
The meeting ends. The reasons should not.