Textbook of AI

The blackboard architecture is a multi-agent AI paradigm introduced in the late 1970s, most famously in the Hearsay-II speech-understanding system (Erman, Hayes-Roth, Lesser and Reddy, 1980). It addresses a fundamental coordination problem: many specialised AI components need to communicate, but a fully connected mesh of point-to-point messages becomes unmanageable.

The metaphor

The architecture is named for the human equivalent: a roomful of experts solving a problem by writing intermediate hypotheses on a shared blackboard. No expert speaks directly to another; instead each watches the board, contributes when they have something relevant to add, and reads what others have written.

Three components cooperate:

1. Knowledge sources (KSs), independent specialists, each with its own representation, methods and triggering conditions. In Hearsay-II these were acoustic, phonetic, lexical, syntactic and semantic modules.
2. The blackboard, a shared global data structure organised into hierarchical levels (e.g. signal → segment → phone → syllable → word → phrase → sentence). KSs read from and write to specific levels.
3. The control / scheduler, picks which KS to run next based on the current blackboard state, KS pre-conditions, and a control strategy. Hearsay-II used an opportunistic agenda; later systems (BB1, GBB) made the control itself a separate blackboard.

Hearsay-II

Hearsay-II (1971–1976), funded by the DARPA Speech Understanding Research programme, recognised connected speech with a 1{,}011-word vocabulary at $\approx 90$% accuracy on a constrained domain. Its modular blackboard design let DARPA's heterogeneous research community contribute components that need not understand each other, only the blackboard schema.

Influence

The blackboard idea anticipated several modern paradigms:

• Publish–subscribe message buses, components communicate by topic on a shared bus rather than directly.
• Tuple spaces, Linda, JavaSpaces, modern actor frameworks.
• Cognitive architectures, ICARUS, CLARION, and parts of SOAR and ACT-R use blackboard-like working memory.
• Robotics middleware, ROS topics behave like a distributed blackboard.
• LLM agent scratchpads, the working-memory regions in tool-using LLMs (ReAct, AutoGPT, multi-agent frameworks) are recognisably blackboards: agents post intermediate hypotheses for other agents and tools to react to.

Successor research (Erman's BB1, Corkill's GBB) refined the framework in the 1980s. The blackboard architecture has had a quiet but persistent influence on robotics, distributed AI and now agentic LLM systems, where a shared scratchpad-as-blackboard mediates collaboration between language model "experts," tool executors and verifiers.