Lesson 8 · Blackboards in modern agent harnesses

The Blackboard, Fifty Years Later

Picture a typical modern multi-agent LLM system: an orchestrator model breaks a task into subtasks; several sub-agents (or tool calls) each work a piece and post results back to a shared context / scratchpad; the orchestrator reads what’s been posted, decides what’s still needed, and dispatches the next round — repeating until the task looks done.

You already have the vocabulary to map this exactly:

  • What on the blackboard-architecture side plays the role of the blackboard itself?
  • What plays the role of knowledge sources?
  • What plays the role of control (the agenda, focus of attention)?
  • What plays the role of a hypothesis (recall: a claim, with a confidence, linked to its support)?

Do that mapping for yourself first — it should feel almost too easy, which is the point: this track has been arguing the correspondence is real, not decorative.

Then the harder question, the one without a ready-made 1970s answer: name one property of the modern setup that Hearsay-II’s designers never had to deal with, and explain what it changes about the design. (Hint: think about what a “hypothesis” is in each system, and who — or what — is doing the matching.)

Work out your answer first — then check it.