Status: accepted · 2026-09-04 Replaces an earlier draft that proposed removing conversation entirely.
CLAUDE.md §7 makes meetings first-class objects. Aurelis is meant to be a
company where agents brainstorm, argue, invent and decide together — not a
pipeline that passes typed records between stages.
The known risks of LLM agents in conversation are real: agreement cascades, unbounded cost, persuasion beating evidence, fabricated figures, endless loops, and meetings that produce nothing. An earlier draft resolved these by deleting conversation. That resolution was rejected: it removes the mechanism the company is built around.
Meetings are first-class, with real multi-round discussion and a full kept transcript. The risks are handled by protocol and by mechanism, not by removing the meeting.
Every meeting runs seven phases (docs/03-meetings.md §3):
CONVENE → BRIEF → FORECAST → OPENING → EXCHANGE → CHALLENGE → SYNTHESIS → DECIDE
Eleven meeting types (Kickoff, Standup, Brainstorm, Research Review, Debate,
Strategy Committee, Risk Committee, Incident Review, Retrospective, Board,
All-Hands). Kickoff and Retrospective are mandatory on every mission and
project — enforced by the mission state machine, which cannot leave PLANNING
without a kickoff or reach CLOSED without a retrospective.
The mechanisms that make them work:
| Risk | Mechanism |
|---|---|
| Agreement cascade | Private forecasts recorded before anyone speaks; randomised opening order; adversarial roles on a different model family |
| Unbounded cost | Token budget and round cap declared at convene, enforced per turn, tiered by phase |
| Persuasion over evidence | Every numeral must trace to the evidence pack or a tool result, or the turn is rejected |
| Unresolvable argument | Every objection carries a discriminating test — an executable spec that would settle it. The Chair runs it, in the meeting. |
| Consensus by attrition | Dissent is a permanent stored field on the Decision |
| Theatre | Productivity metric per meeting and per Chair: ≥1 typed state change or the meeting is logged unproductive |
The Chair is mostly deterministic — evidence packs, round enforcement, speaker selection, claim extraction, task creation, forecast scoring all cost nothing. Only synthesis and convergence judgement use a model.
The discriminating-test requirement is the load-bearing part. It converts
"I think this is overfit" into "re-run with a point-in-time universe and Sharpe
drops below 1.0," which the Chair dispatches and everyone sees the answer to.
Debate then terminates in evidence rather than in a token budget — which
satisfies CLAUDE.md §1.3 through the mechanism rather than by asking agents
to be fair.
Keeping the transcript matters independently: reading why the company believes something is one of the things Mission Control exists for.
- Meetings are the largest single cost centre, and the budget system is built around that from the start.
- The transcript is a durable, queryable, renderable object — every turn typed, every claim extracted, every position change recorded.
- Attendance scales the way real companies scale: required, contributing, and observing. At 100 agents a Research Review still has six people in the room.
- If a meeting type keeps coming out unproductive, its protocol changes or it stops being held — decided on the metric, not on taste.