Overfit by Karl Mayer · Precision that misses the point.
← All posts

We Had Blackboards Once

Two researchers in 1970s Carnegie Mellon laboratory pointing at a blackboard covered in a modern multi-agent system architecture diagram.
The diagram is anachronistic. The problem isn't. Pawn to queen four.

In 1976, a team at Carnegie Mellon built one of the first systems that could recognize continuous human speech. Not isolated words. Not a small, curated vocabulary. Actual speech — messy, ambiguous, full of noise and human variation.

They called it Hearsay‑II. And the challenge it tackled has a name: an ill‑structured problem. In these problems, evidence arrives unevenly, the solution path isn't known in advance, and any architecture that assumes a clean sequence of steps is guaranteed to fail.

The blackboard model didn't fail. It got stranded. When connectionism took hold in the late 1980s and neural networks began outperforming symbolic AI on benchmark after benchmark, the entire paradigm went with it — not because the blackboard model was wrong, but because the field moved and never looked back. We didn't abandon it because it stopped working. We abandoned it because something else worked better on different problems. The distinction matters now, because we're building systems that face the same class of problems Hearsay‑II was designed for, and we're mostly not using the approach that worked.

The room #

Imagine a large whiteboard in the middle of a room. Around it stand specialists: one who understands acoustic signals, one who knows phonemes, one who handles syllables, one who knows vocabulary, one who understands grammar. None of them can solve the problem alone. But all of them can read from and write to the blackboard.

1. The board has levels. A specialist notices something on the board within their expertise and posts a hypothesis — a candidate interpretation with a confidence score. Nobody jumps levels. Acoustic signals at the bottom, then phonemes, syllables, words, phrases, sentences at the top. The hierarchy is explicit and enforced.

2. Hypotheses compete. Another specialist uses that hypothesis as a foothold for the next level up. Multiple candidates coexist at each level until evidence eliminates them. The system avoids premature commitment by design.

3. A scheduler coordinates. Competing interpretations accumulate. Evidence sorts them out. A meta‑level controller decides which specialist acts next based on what's currently most promising — not a fixed pipeline. Opportunistic. The solution path emerges from the evidence, not from a predetermined sequence.

Why the supervisor pattern falls short #

Modern AI systems have already discovered that simple pipelines break on complex tasks. The field's answer has been the supervisor pattern: a central LLM that dispatches to worker agents, interprets their outputs, and decides what to run next — a real improvement that routes dynamically and adapts to what workers return.

But it only recovers one piece of Hearsay‑II — the coordinator — and misses the rest.

The medium is wrong. In the supervisor pattern, workers report to the supervisor, who synthesizes and decides, making the supervisor itself the communication medium. In the blackboard model, agents never talk to each other or to the coordinator directly; they only read and write shared state. When the board is the medium, state is explicit, persistent, and inspectable independently of any agent's behavior. When something goes wrong, you can examine what was asserted, when, and with what confidence. Failures have addresses.

Uncertainty collapses too early. In the supervisor pattern, a worker returns an output and the supervisor either accepts it or reroutes. Either way, the system treats that output as the current best answer. The blackboard model keeps multiple hypotheses alive until evidence eliminates them. It preserves uncertainty instead of flattening it.

There are no abstraction levels. In the supervisor pattern, workers can produce outputs at any level of abstraction and the supervisor handles it. In the blackboard model, levels are architectural constraints — a phoneme‑level specialist cannot produce a sentence‑level hypothesis. The hierarchy is enforced, not conventional.

The supervisor recovered the coordinator, but not the board, the hypotheses, or the levels — which happen to be the parts that matter most for ill-structured problems.

The supervisor pattern answers the question: how do I coordinate multiple agents toward a known goal? It does not answer: how do I solve a problem whose solution path I don't know yet?

The problem that's coming #

Consider a software‑engineering agent working on a large, ambiguous codebase — not a single task but a long‑horizon goal: refactor this system, fix this class of bugs, implement this feature from an underspecified requirement.

An agent is asked to migrate a monolith to a service-oriented architecture. At step 2, it identifies a core abstraction — call it the "user session" — and classifies it as stateless. That classification is plausible. It's also wrong. Steps 3 through 15 build on it: services are scoped around it, interfaces are designed to it, data models derive from it. By step 15, the agent has produced technically coherent code for the wrong architecture. Every individual step was reasonable. The system never had a mechanism to revisit step 2 because step 2 wasn't a hypothesis — it was a fact, routed forward.

Now consider an R&D agent — not executing against a known spec but assembling meaning from contradictory, incomplete evidence where even the right question is uncertain. The work isn't producing a deliverable; it's maintaining a structured set of competing interpretations and letting evidence adjudicate. Collapsing to a single hypothesis early isn't just inefficient — it's the failure mode that defines bad science.

These aren't hypothetical systems. They're being built right now. And in both cases, evidence arrives across multiple levels simultaneously — raw data, intermediate interpretation, theoretical implication, architectural constraint. A supervisor pattern has no way to represent that structure: it can't reopen a higher‑level question when a lower‑level finding changes, hold competing interpretations in parallel, or do anything other than route forward.

What you can build today #

The Hearsay‑II principles translate directly into design decisions you can make.

  • The blackboard becomes a structured shared store (typed, inspectable, append‑friendly) that agents read and write instead of passing context directly. State becomes explicit. Failures become inspectable.
  • Knowledge sources become scoped agents with constrained roles. Not "an agent that does research," but an agent that operates at a specific abstraction level and only produces outputs appropriate to that level.
  • The level hierarchy becomes architectural. You decide upfront what the levels are, what moves between them, and who reads what.
  • Competing hypotheses are preserved. Multiple agents produce candidates at the same level; higher‑level agents evaluate them. Uncertainty survives longer.
  • The scheduler becomes an opportunistic orchestrator that decides what runs next based on the current state of the board.

One honest caveat: the LLM inside each agent doesn't respect your level structure. It will still internally collapse levels in a single forward pass. You're enforcing the structure at the API boundary, which gives you interpretable failure modes and structured handoffs, but not the epistemic guarantees Hearsay‑II's levels provided. Epistemic, here, meaning: what the system is allowed to know, at what level of abstraction, and when. The scaffold is external.

The scaffold being external is what makes it inspectable.

The question #

The field isn't making a mistake. Pipelines at scale solved real problems quickly, the supervisor pattern improved on them, and the architecture we have was shaped by the problems we'd already solved.

The problems we're building toward are different — long-horizon, ambiguous, requiring reasoning across multiple levels of evidence with compounding commitment. The complexity ceiling for agent systems is rising fast, and the dominant failure mode is already visible in systems that are breaking: not hallucination, not capability gaps, but premature commitment routed forward through a coordinator with no shared board to inspect, no competing hypotheses to hold, and no levels to enforce.

There's a reasonable chance we don't ignore this so much as rediscover it — piece by piece, through accumulated pain — and build blackboard-like systems without knowing that's what they are. Which means we'll make the same design decisions Hearsay-II already made, without the fifty years of reasoning about why.

We had blackboards once. We'll have them again — this time, by choice.

— Karl

P.S. One thing this piece doesn't answer: what should actually go on the board. What agents write, what they read, what counts as a hypothesis versus a conclusion, and how you design the vocabulary for each level — none of that is obvious, and getting it wrong undermines everything else. The room is the easy part. That's next.

Further reading

Two recent papers applying blackboard architecture to LLM multi-agent systems — and showing it works: