The mapping:
| Blackboard concept | Modern harness analogue |
|---|---|
| Blackboard | Shared context window / scratchpad / conversation state |
| Knowledge sources | Sub-agents, tool calls, individual model invocations |
| Control (agenda, focus of attention) | The orchestrator’s dispatch loop — deciding which sub-agent or tool runs next |
| Hypothesis (claim + confidence + support links) | A tool result, sub-agent output, or intermediate conclusion posted to context |
Read that table back against lesson 4’s original list and it lines up piece for piece — this track has been arguing that the harness you’ve built already re-derives 1976, and this is the lesson where that claim gets cashed out explicitly rather than gestured at.
What genuinely has no 1970s analogue: the knowledge sources stopped being narrow. Every Hearsay-II KS was a specialist — hand-written, single-purpose, matching one specific pattern (this acoustic KS only ever proposes phonemes; it cannot also parse grammar). A modern sub-agent or tool-calling LLM is a general reasoner that can, in principle, do any KS’s job depending on how it’s prompted — the boundary between “knowledge source” and “control” gets blurry, because the same general model that’s proposing a hypothesis can also be the one deciding what to do about it. Hearsay-II’s designers had to hand-craft dozens of narrow experts because nothing available in 1976 could flexibly do many kinds of reasoning; that constraint doesn’t exist anymore, and its absence changes the design problem from “decompose the domain into experts” to “decide how much to decompose at all” — over-decomposing into needless narrow sub-agents is now a real anti-pattern, where it used to be the only option.
A second consequence worth naming (the hint’s real target): what “confidence” is. Hearsay-II’s hypotheses carried an explicit, engineered numeric confidence that the scheduler (lesson 6) could do arithmetic on — credibility ÷ cost was a real, well-defined computation. A modern tool result or sub-agent output is usually natural language, with no comparably principled confidence attached (a model can be asked to say how sure it is, but that self-report is not the same kind of calibrated, cause-linked number lesson 6’s KSARs carried). Losing that number is a real cost — the elegant priority arithmetic doesn’t transfer cleanly — and it’s exactly why prompt-engineering an orchestrator to explicitly request and weigh confidence, or building an external scoring layer, recreates by hand something the 1976 architecture got for free by construction.
What carries over untouched: the core architectural bet — that opportunistic, evidence-driven control beats a fixed pipeline whenever sub-results need to revise each other — is exactly as true for LLM harnesses as it was for speech. What changed is what’s cheap and what’s expensive: narrow specialists used to be the only option and are now optional; principled numeric confidence used to be free and is now the part you have to engineer back in.
Where this goes: Stage 1 built the blackboard payoff from the ground up — anatomy, control, when it’s warranted, where it landed today. Stage 2 goes back a level, into the integrated cognitive architectures (ACT-R, SOAR) that took production systems and grew them into full models of a mind — starting with ACT-R’s split between what you know and how you act on it.