The Blackboard Is Back: What 1980s AI Knew About Multi-Agent Coordination

If your agent team coordinates through a shared plan file, a repo, or a design doc that everyone reads and writes, congratulations: you have reinvented the blackboard architecture. It was state of the art in 1975. The uncomfortable part is not the reinvention — good ideas deserve to come back. The uncomfortable part is that the original had three load-bearing components, and most modern agent stacks rebuilt only one of them.

Hearsay-II, the DARPA-funded speech understanding system built at Carnegie Mellon between 1971 and 1976, faced a problem that should sound familiar: many specialized, unreliable experts — acoustic analyzers, syntax predictors, semantic raters — none of which could solve the problem alone, all of which needed to build on each other's partial guesses. The architecture that emerged had a shared workspace (the blackboard), independent specialists (knowledge sources), and a scheduler that decided, at every step, which specialist's contribution was worth executing next. Fifty years later, teams wiring LLM agents together are converging on the same shape — a lead agent, a set of workers, a shared artifact — and hitting failure modes the blackboard literature named and solved before most of us were born.

The architecture, and the two-thirds we forgot

A classical blackboard system has exactly three parts, and the discipline lives in how strictly they are separated.

The blackboard is a shared, structured data store where all partial solutions live. Not a chat log — a structured hierarchy. In Hearsay-II, hypotheses lived at explicit levels (acoustic segments, syllables, words, phrases), each entry carried a credibility rating, and links recorded which lower-level evidence supported which higher-level guess. Any specialist could see the current best picture of the whole problem at any time.

The knowledge sources are independent specialists that never talk to each other. This is the rule modern builders find hardest to swallow: knowledge sources communicate only through the blackboard. Each one is a condition–action pair — "if hypotheses of this shape appear, I can contribute" — with no knowledge of which other specialists exist. Adding a new specialist requires changing nothing else in the system.

The control mechanism is a scheduler that watches the blackboard and decides which activated specialist actually runs next. This is the part everyone forgets, and it was the part the original researchers considered the actual research contribution. Hearsay-II's scheduler estimated, for every candidate action, how much it would improve the emerging solution, and executed the most promising one — a strategy the literature calls opportunistic scheduling. The system didn't run top-down or bottom-up; it ran wherever the current evidence was strongest, island-hopping from confident words outward into ambiguous audio.

Later systems went further. Barbara Hayes-Roth's BB1 architecture at Stanford (1984) put the control decisions themselves on a second blackboard, so the system could reason about its own strategy — construct an explicit plan of attack, explain why it chose each action, and revise the strategy mid-run when it stopped paying off. Read that again in 2026 terms: an agent system whose orchestration policy was itself inspectable, explainable state, not vibes buried in a prompt.

Modern agent teams that coordinate through a shared artifact have rebuilt the blackboard. Very few have rebuilt the credibility ratings, the level structure, or the scheduler. We kept the whiteboard and fired the meeting facilitator.

Why message-passing swarms fail in ways a blackboard can't

The dominant alternative to a shared workspace is direct agent-to-agent messaging: orchestrators DM-ing workers, workers handing results to other workers, "society of agents" designs where coordination lives in the message stream. The empirical record on this is now substantial, and it is grim. The MAST study — the largest failure taxonomy for LLM multi-agent systems to date, built from over 1,600 annotated execution traces across seven popular frameworks — found that roughly 37% of all failures are inter-agent misalignment: agents withholding or losing information across handoffs, ignoring each other's inputs, or proceeding on incompatible assumptions. Add the specification failures that shared visible state would have surfaced, and well over half the observed failures are coordination failures, not capability failures.

The mechanism is not mysterious. Every agent-to-agent message is a lossy re-summarization performed by a model under token pressure. Chain three handoffs and you are playing telephone with your own architecture. The blackboard makes an entire class of these failures structurally impossible rather than merely less likely:

• Context loss during handoff can't happen, because there is no handoff. The partial solution stays on the blackboard in full fidelity; the next specialist reads the original, not a summary of a summary.
• Conflicting parallel work stays visible. When two agents post incompatible hypotheses to a shared workspace, the conflict is sitting in public state where the scheduler — or a human — can see and resolve it. In a message-passing design, each conflicting belief lives privately in a different context window, and you discover the disagreement at merge time, as a mystery diff.
• "Who needs to know this?" disappears as a routing decision. A messaging agent must decide who to inform, and models are bad at that decision under pressure. On a blackboard, publishing and routing are the same act.

Loading…