Experiment

Multi-Agent Coordination: Sequential State-Passing vs a Shared Blackboard

Compare the custom runtime's current coordination model, where agents run in rounds and pass a handoff envelope, with a blackboard model where agents read and write a shared workspace, on convergence, token use, cost and latency.

DESIGNED · NOT RUN YET

Hypothesis

A shared blackboard should cut contradictions before reconciliation, because each agent sees others' commitments as they are made, but it will raise token use, because every agent reads more context. Sequential state-passing should stay cheaper and more predictable.

Problem

In the executive-team run, specialists build on each other in fixed rounds, and the CEO agent finds and resolves conflicts at the end. That works, but conflicts are only caught late, and every revision round costs extra calls.

Setup

Baseline: the current seven-round run. The CEO writes sub-briefs, the CMO sets the foundation, Creative and Sales work in parallel on it, the PM builds the roadmap, the CEO scans for conflicts, flagged agents revise, and the CEO writes one strategy. State moves only through the handoff envelope. Variant: a blackboard of typed entries (positioning, price points, channels, dates, risks) that every specialist reads before writing and appends to after. The CEO reconciles at the end as before. Metrics: conflicts flagged by the CEO's scan, revision rounds needed, total tokens, LLM calls, cost per run and wall-clock time.

DIAGRAM

Two ways for agents to coordinate

flowchart TD
  subgraph SP[Sequential state-passing]
    C1[CEO sub-briefs] --> M1[CMO foundation]
    M1 --> X1[Creative + Sales]
    X1 --> P1[PM roadmap]
    P1 --> K1{CEO conflict scan}
    K1 --> R1[Revisions]
    R1 --> S1([Strategy])
  end
  subgraph BB[Shared blackboard]
    B[(Blackboard: positioning, prices, channels, dates, risks)]
    M2[CMO] <--> B
    X2[Creative] <--> B
    Y2[Sales] <--> B
    P2[PM] <--> B
    B --> K2{CEO reconcile}
    K2 --> S2([Strategy])
  end

In state-passing, each round hands an envelope to the next and conflicts are found at the end. With a blackboard, every specialist reads and writes one shared workspace as it goes.

Input

A fixed set of representative briefs (launch plans, campaigns, event decisions), run several times through each variant with the same models and temperatures.

Output

Per-variant averages for each metric, and examples of conflicts each model caught or missed.

Decision

Status: designed, not run. The baseline runtime is in production; the blackboard variant is not built yet.

Next Step

Implement the blackboard as an alternative workflow behind a setting, run the brief set through both, and publish the comparison here.

Use Cases

▸Choosing a coordination model for multi-agent systems
▸Reducing late-stage conflict resolution cost