Skip to content

Latest commit

 

History

History
332 lines (264 loc) · 16.8 KB

File metadata and controls

332 lines (264 loc) · 16.8 KB

03 — Meetings & Collaboration

Date: 2026-09-04 Status: proposal, v2.

Meetings are how this company thinks together. Agents brainstorm, argue, present evidence, change each other's minds, and decide. The transcript is kept — it is part of the company's record, and reading it is one of the things Mission Control is for.

This document specifies how a meeting works so that it produces real work instead of expensive theatre.


1. The rule that makes meetings worth having

Every meeting ends in state changes, and the state changes are typed records — not a summary paragraph.

A meeting that produces no task, no decision, no experiment, no revision and no recorded disagreement is logged as unproductive, and that number is a metric on the Chair and on the meeting type. Meeting types that keep coming out unproductive get their protocol changed or get stopped.

That single rule is what separates this from roleplay. Everything below is in service of it.


2. Meeting types

Type Trigger Participants Rounds Produces
Kickoff Every mission and project starts Assigned team + department heads 5 Plan, hypothesis backlog, task assignments, budget split, forecasts
Standup Scheduled, cheap One team 1 Status deltas, blockers, reassignments
Brainstorm Research Director or Head of Strategy calls it Cross-department, deliberately mixed 4 Candidate hypotheses, leads, research directions
Research Review A finding is ready or contested Author, Critic, Statistician, Auditor, Lead 6 Verdict on the finding, objections with tests, follow-ups
Debate Unresolved disagreement or a contested claim Two sides + Chair + neutral Auditor 6 Position record, discriminating tests, decision, preserved dissent
Strategy Committee A strategy requests promotion Head of Strategy, Validation, Risk, Audit, CIO 6 Promotion decision against the gates, or the specific reason it failed
Risk Committee Limit change, exposure breach, new deployment CRO, Risk, PM, Exposure, Correlation, Head of Trading 4 Limits, vetoes, mandated reductions
Incident Review Alert, breach, data failure, live divergence Whoever is implicated + Audit + Ops 4 Root cause, corrective tasks, standing rule
Retrospective Every mission and project ends Everyone who worked on it 4 Lessons to memory, org change proposals, calibration review
Board Weekly, plus escalations and org changes Executive + all department heads 6 Mission portfolio, priorities, org changes, budget
All-Hands Monthly or on a major result Everyone 3 Company state, direction, recognition of good failures

Kickoff and Retrospective are mandatory on every mission and project. That is the "meet at the start and meet at the end" pattern — enforced by the mission state machine, which cannot leave PLANNING without a Kickoff and cannot reach CLOSED without a Retrospective.

Everything between them runs on individual work, messages and channels, with the heavier meeting types called when they are needed.


3. The meeting protocol

Every meeting runs the same seven phases. Types differ in which phases are enabled, how many exchange rounds they get, and who may speak in each.

┌─ 0. CONVENE ────────────────────────────────────────────────────────────┐
│  Deterministic. The Chair assembles:                                     │
│    · agenda (from the trigger + open action items)                       │
│    · EVIDENCE PACK — every relevant artifact, finding, metric and prior   │
│      decision, pre-fetched with hashes                                    │
│    · participant list, round cap, token budget                           │
│  Costs nothing. No model call.                                           │
├─ 1. BRIEF ──────────────────────────────────────────────────────────────┤
│  Deterministic. The state of the world, rendered from the record:        │
│  what is known, what is being asked, what was decided before, what       │
│  failed before. Agents read the same brief — no information asymmetry     │
│  by accident.                                                            │
├─ 2. FORECAST ───────────────────────────────────────────────────────────┤
│  Each participant privately records a probability for the outcome,        │
│  BEFORE hearing anyone. Cheap (low tier). Scored later. This is the      │
│  company's honest quality signal and it costs almost nothing.            │
├─ 3. OPENING ────────────────────────────────────────────────────────────┤
│  One bounded turn each. Position, key evidence refs, what would change   │
│  their mind. Everyone speaks exactly once.                               │
├─ 4. EXCHANGE ───────────────────────────────────────────────────────────┤
│  The actual discussion. N rounds, capped. Participants see each other's  │
│  turns and respond. Brainstorming, challenge, building on ideas.         │
│  The Chair selects who speaks each round — normally those in genuine     │
│  disagreement, plus anyone directly asked a question.                    │
├─ 5. CHALLENGE ──────────────────────────────────────────────────────────┤
│  Objections are formalised. Each must carry a DISCRIMINATING TEST:       │
│  an executable spec that would settle it. The Chair dispatches every     │
│  test that fits the meeting's compute budget; results come back into     │
│  the room if they finish in time, otherwise as follow-up tasks.          │
├─ 6. SYNTHESIS ──────────────────────────────────────────────────────────┤
│  The Chair (or the senior decision-maker for this type) drafts the       │
│  outcome: what was agreed, what remains contested, what is decided.      │
├─ 7. DECIDE & ASSIGN ────────────────────────────────────────────────────┤
│  Typed Decision record with DISSENT PRESERVED — who disagreed, why, and  │
│  what evidence they cited. Action items become real Task rows with       │
│  owners and deadlines. Forecasts are scored against the outcome.         │
└──────────────────────────────────────────────────────────────────────────┘

Why phase 5 has teeth

An objection that must carry a test that would settle it is the difference between a meeting and an argument. "I'm worried this is overfit" is not an objection. "Re-run this with a point-in-time universe and the Sharpe will drop below 1.0" is — and the Chair runs it, in the meeting, and everyone sees the answer.

That mechanism is why the debate ends in evidence instead of ending when someone gets tired.


4. What a turn looks like

Every contribution is a typed artifact, which is what makes the transcript queryable and the meeting auditable.

Turn:
    turn_id · meeting_id · round · phase
    speaker_agent · addressed_to[]
    kind      POSITION | ARGUMENT | QUESTION | ANSWER | EVIDENCE
              | OBJECTION | CONCESSION | PROPOSAL | SYNTHESIS
    body                    natural language — this is a real conversation
    claims[]                each factual claim, separately extracted
    evidence_refs[]         artifact hashes / finding ids per claim
    stance                  SUPPORTS | OPPOSES | ABSTAINS | UNCERTAIN
    changed_mind_from       set when a participant moves position
    tokens · cost

Two validators run on every turn:

  1. Numbers must be sourced. A figure in a turn body that does not appear in the evidence pack or in a tool result from this meeting is rejected. Agents quote the record; they do not recall numbers.
  2. Claims must be typed. A factual claim without an evidence ref must be explicitly marked as opinion, speculation or a question. This is not censorship — speculation is genuinely useful in a Brainstorm, and marking it is what keeps it from hardening into a fact three meetings later.

changed_mind_from is deliberately tracked. An agent that updates on evidence is doing the job; an agent that never updates, or always updates, is a measured problem.


5. The Chair

The Chair is partly deterministic software and partly the Chief of Staff agent.

Deterministic (no model call):

  • assembling the evidence pack and the brief
  • enforcing round caps, token budgets and turn limits
  • selecting speakers by stance conflict and by direct address
  • dispatching discriminating tests
  • extracting claims, evidence refs and stances from turns
  • creating Task rows from action items
  • scoring forecasts
  • computing the productivity metric

Model-driven (Chief of Staff, mid tier):

  • drafting the synthesis
  • deciding when a discussion has converged early
  • deciding when to escalate an unresolved disagreement to a higher meeting

Splitting it this way means most of the Chair costs nothing, and the parts that cost money are the parts that need judgement.


6. Cost control

Meetings are the most expensive thing the company does. They are budgeted like anything else.

Per-meeting budget

Declared at CONVENE, enforced by the Chair:

MeetingBudget:
    max_participants        typically 4–8; All-Hands is the exception
    max_exchange_rounds     3–6 by type
    max_tokens_total        hard cap
    max_tokens_per_turn     keeps one agent from consuming the room
    max_test_dispatches     compute budget for discriminating tests
    tier_policy             who gets which model in which phase

When the budget is exhausted, the Chair moves straight to SYNTHESIS and the unfinished threads become follow-up tasks. Running out of meeting is a normal outcome, not a failure.

The cost ladder

Phase Tier Why
CONVENE, BRIEF none Deterministic
FORECAST low One probability each
OPENING mid, low for junior One bounded turn each
EXCHANGE mid; high for the two principals in a Debate Where the thinking happens
CHALLENGE mid Objections are structured
SYNTHESIS mid One draft
DECIDE none Typed extraction

Attendance discipline

Not everyone attends everything. Participation is resolved from the subject:

  • Required — anyone whose write scope covers the decision (Risk for a risk decision, Head of Strategy for a promotion).
  • Contributing — those with relevant evidence or genuine stance conflict.
  • Observing — receives the minutes, does not speak, costs nothing.

At Stage 5 with a hundred agents, a Research Review still has six people in the room and ninety-four reading the minutes. That is how real companies scale meetings and it is how this one does too.


7. Minutes and the record

Every meeting produces:

Meeting:
    meeting_id · type · subject · trigger
    mission_id · project_id · desk · department
    chair · participants[] (with roles at the time)
    agenda[] · evidence_pack[] (hashes)
    turns[]                      the full transcript
    forecasts[] · forecast_scores[]
    objections[] · discriminating_tests[] · test_results[]
    decisions[] · dissent[]
    action_items[] → Task rows
    tokens · cost · duration
    productive: bool             ≥1 state change

The transcript is kept in full and is readable in Mission Control. Clicking a meeting shows the room, who said what, who changed their mind, what evidence was cited, what was decided, and who disagreed.

The minutes are extracted deterministically from the turns, so the summary cannot disagree with the transcript.


8. How meetings connect to the work

MISSION OPENS
    │
    ├─▶ KICKOFF ────────────────────▶ plan · hypotheses · tasks · forecasts
    │                                          │
    │        individual work, messages, channel posts, engine runs
    │                                          │
    ├─▶ BRAINSTORM (as needed) ─────▶ new hypotheses, new directions
    ├─▶ STANDUP (scheduled) ────────▶ blockers, reassignment
    ├─▶ RESEARCH REVIEW (per finding) ▶ verdict, objections, follow-ups
    ├─▶ DEBATE (on deadlock) ───────▶ discriminating tests, decision, dissent
    ├─▶ STRATEGY COMMITTEE ─────────▶ promotion or specific refusal
    ├─▶ RISK COMMITTEE ─────────────▶ limits, vetoes
    │                                          │
    └─▶ RETROSPECTIVE ──────────────▶ lessons → memory
                                      calibration review
                                      org change proposals → Board
MISSION CLOSES

The Retrospective is where the company learns. It reviews what was predicted against what happened, extracts lessons into institutional memory as standing rules, and — critically — feeds the Org Development Lead with the evidence for structural change proposals.


9. Brainstorming, specifically

The user asked for a company that invents things. Brainstorm meetings are built for divergence, and they run under different rules from the rest:

  • Speculation is allowed and encouraged, marked as such. Nothing said in a Brainstorm can become evidence for anything; ideas leave as LEAD nodes in the knowledge graph, never as findings.
  • Deliberately mixed attendance. A Brainstorm on options volatility pulls in the crypto Technical Analyst and the Macro Analyst on purpose — the cross-desk transfer is the point.
  • No decision required. A Brainstorm's productive output is candidate hypotheses, and it is productive if it produces any.
  • Prior art is pre-loaded. The Knowledge department injects "we tried this before, here is what happened" into the evidence pack automatically. This is the single most valuable thing institutional memory does — 174 trials of graveyard is exactly the context a brainstorm needs to not repeat itself.
  • Novelty scoring. Candidate hypotheses are checked against the ledger for duplication before they consume research budget.

10. Failure modes, and what stops them

Meetings between LLM agents have known ways of going wrong. Each gets a mechanism, not a warning.

Failure Mechanism
Agreement cascade — everyone agrees with the first speaker Private forecasts recorded in phase 2 before anyone speaks; opening turn order randomised; adversarial roles pinned to a different model family
Endless discussion Hard round and token caps; the Chair moves to synthesis on exhaustion
Persuasion beating evidence Numbers must be sourced; objections must carry tests; the test result decides
Fabricated figures Turn validator rejects unsourced numerals
Consensus by attrition Dissent is a stored field on the Decision, permanently
Roleplay with no output Productivity metric per meeting and per Chair; unproductive types get changed or stopped
Cost blowout Budget declared at convene, enforced per turn, tiered by phase
The loudest agent dominates Per-turn token cap; the Chair selects speakers by stance conflict, not by eagerness
Meetings for their own sake Only Kickoff and Retrospective are mandatory; everything else needs a trigger
Same debate every month Prior decisions are in the evidence pack; re-opening a settled question requires a stated new reason

11. What this gives you

A company where agents genuinely work together: they observe, they bring findings to each other, they argue, they run tests to settle arguments, they change their minds on evidence, they decide, they disagree on the record, and they go away with assignments.

And every one of those conversations is a durable, queryable, auditable object that Mission Control can render — so you can open a meeting from three weeks ago and see exactly why the company believes what it believes.