Blackboard Pattern

Independent components cooperating through a shared knowledge store.

Also known as: blackboard pattern, blackboard architecture, blackboard system

Blackboard is an architecture where several independent specialists cooperate through a shared data store — the blackboard. Each specialist (a “knowledge source”) watches the shared state, contributes when its expertise applies, and the combination of partial contributions gradually builds toward a solution. A control component decides which specialist acts next.

knowledge sources:  parser, predictor, detector, ...
                        │ read/write
blackboard:        [ shared evolving state / partial solutions ]
                        │
control:           decides who runs next

The metaphor is a group of experts at a blackboard: each adds what they know, sees what others have written, and the answer emerges from the collaboration. It fits problems that are too varied for a single fixed algorithm — the system doesn’t know in advance which combination of techniques will crack a given case.

Classic uses are AI-flavoured: speech recognition (acoustic, lexical and language models contributing incrementally), image understanding, planning. Outside AI, it appears as shared-state architectures where many agents contribute to a common model over time.

The classic mistakes:

Confusing it with pub/sub. Pub/sub is a decoupled event broadcast; the blackboard is a shared evolving model that components read and update, with control logic deciding who acts. An event bus tells you something happened; a blackboard holds the current partial answer.
No control strategy. Without a rule for which source runs next, you get either thrashing (everyone reacts to everything) or nothing progressing. The controller is essential, not optional.
Unbounded blackboard growth. The shared state grows as contributions accumulate; without pruning or a termination condition, it keeps growing and the search never ends.
No termination criteria. Because many specialists can keep refining, you need an explicit “good enough” or “no more progress” rule, or the system doesn’t stop.
Reaching for it over-engineering. Blackboard is heavyweight. For most problems a straightforward pipeline or a mediator coordinating a few components is simpler and adequate.

When to use it: when the problem is naturally solved by heterogeneous specialists contributing to a shared solution that evolves — and no single algorithm fits all cases. It’s a niche but powerful architecture, and it teaches a useful idea: sometimes the answer is a shared model plus contributors, not one monolithic solver. It sits near domain-driven design in spirit when a complex domain needs several models contributing to one understanding.