Mastering AI Agent Architectures: Shared Memory: Blackboards, Tuple Spaces, CRDTs, and Consensus: 5 (Mastering AI Agent Architectures — The Styles Series) - Softcover

Inhaltsangabe

Agents that build on each other’s work need a place to put it — and that place can fail, go stale, or conflict.

This book measures eighteen shared-memory styles on the Gauntlet benchmark. It starts from the classic blackboard — knowledge sources, partial entries, an agenda written as code — and then asks what it costs to keep a shared board alive and true.

• The classic blackboard: 96 % on research at 0.79× the cost of a voting board
• Tuple spaces — a hierarchy without a master
• The consistency menu — why stale reads are cheap and strong reads cost the deadline
• CRDTs — convergent is not correct: a merge that applies an adjustment twice
• Consensus only where invariants demand it
• Context coherence — typed views per agent instead of the whole board
• Contention and hot spots — one lock, shards, or optimistic writes
• Shared memory versus messages, measured pair by pair

The decision card is written as seven conditions, each delivered by a chapter, with a production checklist beside it.

Runnable code, frozen results and a lock at tag gauntlet-b5. Book 5 of 7 in Mastering AI Agent Architectures — The Styles Series.

Die Inhaltsangabe kann sich auf eine andere Ausgabe dieses Titels beziehen.