Skip to content
sync ai agents

What is a blackboard in multi-agent AI?

Updated November 2026

A blackboard is a shared working space where agents publish partial findings and inspect information contributed by others. It supports indirect coordination through evolving task state rather than requiring every agent to address every other agent directly.

How does blackboard coordination work?

Microsoft’s architecture guide explains several ways to structure cooperating agents. A blackboard is a shared-state design option rather than a required feature of those patterns. Its defining property is that participants coordinate through a common representation of the problem and intermediate results.

For a diagnostic task, one agent might post observed symptoms, another propose hypotheses, and another attach test results. Entries should distinguish observations, interpretations, proposed actions, and accepted conclusions. Otherwise a speculative hypothesis can become an apparent fact simply because it remains visible.

How is a blackboard different from shared memory?

Shared memory is a broad storage concept. A blackboard specifically organizes active problem solving around shared working state. It may use persistent storage, but persistence alone does not provide coordination. Define entry types, allowed updates, and the conditions under which an agent should react.

ElementUseful content
ObservationEvidence, source, and timestamp
HypothesisProposed explanation and confidence limits
DecisionAccepted outcome and decision owner

How should conflicting contributions be handled?

Avoid treating the most recent write as automatically correct. Preserve provenance, use version checks for concurrent changes, and separate proposals from accepted state. Where agents disagree, keep both claims until a designated rule or reviewer resolves the conflict.

The tradeoff is shared visibility versus contention and information overload. A growing board can increase retrieval cost, expose irrelevant information, and encourage agents to respond repeatedly to the same entry. Use scoped views, task identifiers, and explicit completion markers.

For sensitive work, apply access control to individual tasks or entry classes rather than assuming every participant should read everything. Maintain an update history so a mistaken contribution can be traced and corrected. A blackboard is not itself a scheduler, permission system, or guarantee that agents will converge on a valid answer.

What else should you know?

  • Is a blackboard just a chat log?

    No. Structured working state and update rules distinguish the pattern.

  • Does it require a central supervisor?

    No, although a supervisor or scheduler can govern access and progress.

  • Should agents overwrite each other’s findings?

    Prefer versioned proposals and explicit resolution over silent replacement.

  • When is it useful?

    Use it when specialists iteratively refine a shared problem representation.

Where can you go deeper?