Multi-Agent Wiki
Information

Blackboard / Shared Memory / Workspace

Agents collaborate indirectly through shared state, knowledge, a task board, or a workspace.

Live visualizationAnimated topology — press Space to play / pause
·Diagram
Live
1 / 7

—

Definition

Multiple agents collaborate indirectly through shared state, knowledge base, task board, or workspace.

Category: Information flow

When to use

Long-running tasks, asynchronous collaboration, multi-party shared evidence, code workspaces, task-state management.

When not to use

When the shared state has no versioning, no permissions, and no expiry — it will rot.

How to implement

  1. Partition the blackboard: facts / hypotheses / tasks / artifacts / decisions.
  2. Every record carries source, timestamp, confidence, owner, TTL.
  3. Agents read/write only through APIs — no direct global state mutation.
  4. Surface conflicting facts as a conflict set; do not silently overwrite.

Minimal pseudocode

TypeScript
type BlackboardItem = {
  id: string;
  type: "fact" | "hypothesis" | "artifact" | "decision" | "task";
  content: unknown;
  sourceAgent: string;
  confidence?: number;
  createdAt: string;
  ttl?: number;
};

Recommended trace events

  • blackboard.item.created
  • blackboard.item.updated
  • blackboard.conflict.detected
  • blackboard.item.expired

Common failure modes

  • State pollution.
  • Stale entries treated as fresh.
  • Concurrent writes from multiple agents.
  • No provenance.

Implementation checklist

  • Input/output schemas defined.
  • Each agent's permission boundary defined.
  • Every agent call carries a run id / trace id.
  • Failure, timeout, cancel, and retry strategies defined.
  • Context passed is the minimum required, not the full history.
  • High-risk actions are gated by approval or a verifier.

References