Agent teams & collaboration rooms

Some work is too big or too parallel for one context window. Sonderr’s answer is the collaboration room: one lead agent with write access, joined by read-only AI workers who investigate, review and argue — inside your chat, in the open. This page covers how rooms are built, what the workers can and cannot do, and how a room ends with something you can trust.

1 · The shape of a room

A room is always one lead + workers. The lead is the only agent that can change files, run tools that mutate anything, or act on the room’s conclusions. Workers are strictly read-only: they inspect the workspace, run supplied read-only tools, and report. Since Swarm 2.0 (v1.6) a room supports six workers drawn from sixteen personas across eight roles — Builders, Testers, Documenters and Strategists joined the original roster — and saved rooms restore their full-sized rosters when you reopen them.

Personas are exactly what the UI says they are: labeled fictional AI characters. A worker named “Test” is not a second model with special powers; it’s a role with a brief. That honesty matters for everything else on this page.

2 · A room in motion

Every ceiling in a room is explicit, and generous enough to be useful:

• Six shared findings per worker — concrete observations with file references, not vibes.
• Four risk flags per worker, sixteen per room — things that look wrong, dangerous or missing.
• Twelve live help threads — workers can ask the room questions mid-investigation.
• A 160-entry transcript and two anonymous helper calls per room for the lead to consult a second opinion without breaking the scene.

Findings and flags stream into the chat as they happen — you watch the investigation, you don’t wait for a report at the end. Everything lands in the same live task card machinery as normal work, so a room running inside a Build task keeps its evidence attached to the todo item it justifies.

3 · The synthesis pass

Rooms don’t trail off. Since v1.6, the AI lead ends every room with a published synthesis: a structured verdict that separates agreement (what multiple workers independently saw) from contradictions (where reports conflict) and gaps (what nobody could establish), followed by up to three recommended next steps. The synthesis renders with its own LEAD SYNTHESIS styling so you can’t mistake it for another chat message, and a failed synthesis never blocks the room’s result — the findings stand even if the verdict generation hiccups.

Treat the synthesis as what it is: an AI assessment. It’s the best-organized summary of the evidence, not a permission slip — the same rule that governs every agent output in Sonderr.

4 · Saved rooms and recovery

Rooms are persistent artifacts. Save one and its roster, purpose and findings restore intact when you reopen it — useful for standing review crews (the “security room” that looks at every major change) or for an investigation you’ll return to across days. Recovery guards the usual failure modes: reopening a room can’t be corrupted by a late result from an older view, and every live room snapshot carries the owning conversation’s ID so delayed worker updates can’t be attributed to the wrong chat.

5 · The honesty contract

Rooms are engineered so that enthusiasm can’t become authority:

• Worker identities are labeled fictional; their assessments are unverified evidence until the lead — or you — verifies them.
• Workers cannot authorize anything. No worker report, however confident, grants the lead a new permission; the lead’s actions obey the same task-mode and access rules as ever.
• The lead, never a worker, performs all edits and checks. A room changes your workspace through exactly the hands that own the plan.
• Team-room snapshots stay isolated per conversation, so one chat’s swarm can’t contaminate another’s composer state.

6 · Room, or no room?

Rooms cost tokens and time; they earn it on breadth. Good fits: multi-file investigations, “find everything wrong with this before we ship,” design-space exploration where independent perspectives genuinely diverge, and review of large diffs. Bad fits: focused single-file edits, quick factual questions, anything where you already know the next concrete step — there the lead alone, or a plain Ask, is faster and just as good. When you ask Sonderr to “spawn three agents” or “research this with agents,” it recognizes the request and routes it; when a request is underspecified, it asks rather than burning a room on a guess.