One Memory All Agents Can Read and Write (Blackboards)

Why this matters

Some multi-agent systems coordinate through a shared memory or 'blackboard' — a common store agents read from and write to. It's powerful for loosely-coupled collaboration but dangerous for context: an unbounded blackboard becomes a dumping ground that every agent's window must absorb. Managing a blackboard is context engineering at the team level: structure it, scope reads, bound writes, and compress it over time. Knowing when a blackboard helps (genuinely shared evolving state) vs. when explicit hand-offs are cleaner is a key architectural judgment.

Demo

The demo implements a bounded, structured blackboard: agents post structured entries, readers query by relevance/tag rather than reading everything, and a compaction step keeps it from growing without limit.

Try it yourself

1. Read with a tag filter and confirm an agent pulls only relevant entries, not the whole board.
2. Post past max_entries and watch the board stay bounded instead of growing forever.
3. Replace hard truncation with summarizing the oldest entries into a digest entry.
4. Decide for a real workflow: blackboard (shared evolving state) vs. explicit hand-offs (linear pipeline) — and justify it.

References

• Blackboard system
• Anthropic — Building effective agents

Chat about this lesson