Blackboard Pattern

Picture a team working a hard problem in one room with a whiteboard. Nobody hands out tasks. Each person watches the board and, when they see something they can improve, steps up and writes. That is the blackboard pattern — an architecture that came out of 1970s speech-understanding research, where no single component could solve the problem alone.

The three parts
The separation between them is the whole idea:

The blackboard — a shared, structured store holding the current partial solution: hypotheses, intermediate results, open questions.
The knowledge sources — independent specialists that read the board, apply their expertise, and post contributions back. They never address each other directly.
The control component — which decides who acts next, based on what the board looks like now.

Why agent systems reach for it
Point-to-point messaging between agents gets expensive: every pair needs a channel and every message restates context. A blackboard replaces that with one piece of shared state — a researcher posts findings, a critic posts objections, a writer picks up what survived, and none of them needs to know who else exists. Adding a fourth specialist changes no existing code. LangGraph's shared state object, read and updated by each node, is this pattern in modern dress.

Where it strains
Control is the weak point: with no clear policy for who acts next, agents overwrite each other's conclusions, re-derive what is already posted, or stall because none of them claims the current state. The board grows, too — once it outgrows the context window it needs retrieval and compaction of its own.

The blackboard pattern coordinates multiple agents through a shared store of partial results rather than direct messages: each agent reads the board, contributes what its specialty allows, and a control component decides who acts next. It keeps agents decoupled — a new specialist can be added without rewiring anyone — and concentrates the coordination problem in one place: deciding who writes next, and stopping agents from overwriting or re-deriving each other's work.

Think of a detective squad with one big whiteboard. Nobody is assigned anything. People walk up, read what's there, and add what they can — a name, a timeline, a contradiction. Slowly the case comes together.

The blackboard pattern is that arrangement for AI assistants working as a group. Instead of messaging each other, they all read from and write to one shared noticeboard, each adding what it is good at.

It keeps things simple: you can bring in a new specialist without rewiring anything. The hard part is traffic control — deciding whose turn it is, so two helpers don't rub out each other's work.