The blackboard architecture — born in 1970s speech-recognition systems like Hearsay-II — gives a set of independent specialists a single shared workspace they all read from and write to, incrementally solving problems no single specialist could solve alone.
The design has three parts. A blackboard: global, structured state holding the emerging solution. Knowledge sources: independent specialists that watch the board and contribute when they see conditions they can act on. A control shell: the scheduler that decides which contributor runs next. Nothing messages anything directly — coordination happens entirely through the state of the board, which makes the whole system's progress inspectable at a glance.
Modern agent systems are rediscovering this shape. A coding agent's shared workspace — plan file, notes, task tracker, test results — is a blackboard in everything but name: the planner, coder, and verifier are knowledge sources that opportunistically update it. The pattern shines exactly where agents shine: problems with no fixed solution path, where the next useful step depends on what has been discovered so far, and where different kinds of expertise must combine opportunistically.
Why It Maps So Well to LLM Agents
- Context budgets: each knowledge source reads only the board slice it needs, instead of one agent strangling on a monopoly context window.
- Heterogeneity: a cheap fast model can handle formatting contributions while a deep-reasoning model handles architecture — the board does not care.
- Incrementalism: partial solutions are visible early; human reviewers can inspect a run mid-flight rather than waiting for a monolithic answer.
- Recoverability: because state lives on the board, a crashed contributor loses nothing that matters.
The historical caveats transfer too: blackboards need conflict resolution (two sources editing the same region), termination criteria (who calls the problem solved?), and scheduling discipline — otherwise the loudest specialist, not the most relevant one, drives the run. Those three controls are precisely what separates a blackboard architecture from a shared folder that agents argue over.