Fan-out reads, single-writer state

Ten agents processing the same batch in parallel. How do you stop them corrupting shared state?

The answer we landed on is old: fan out reads, single-writer for state changes.

Before the fan-out even runs, a Redis-only batching layer collects incoming messages until a size threshold or timeout fires. That step is why ten parallel model calls are affordable at all — you're paying one call per agent per batch, not per message. The story usually gets told starting at "ten agents read the same batch," which quietly skips the part that makes the economics work.

Seven model-backed agents — sentiment, embedding at 384 dimensions, toxicity, spam, emoji, topic clustering at a 70% similarity threshold, relationship inference — each make exactly one model call per batch and return a result. Stateless, so they parallelize perfectly, and an eighth costs nothing conceptually.

Three more agents split the rest: one merges every result into the shared graph, one maintains the vector index, one assembles context for the query layer downstream. Only the first writes. Everyone else reads, or stays out of the graph entirely.

Single writer. No locks, no partial merges, no ordering bugs at 3am. The graph is scoped to one batch and discarded once persisted, so a bad batch can't poison the next one, and memory stays bounded no matter how long the pipeline runs.

The persistence step behind the writer is three components, not one: a delta detector doing content-hash idempotency, a normalized entity writer, and a serializer that snapshots the whole graph as a blob. Normalized rows for querying, a snapshot for replay — the same state written twice, on purpose.

The instinct with agent systems is to let each own its slice of state. It feels more autonomous. It also means every new agent multiplies the interleavings you have to reason about, and failures show up as corrupted relationships at 3am, not a stack trace. The cost of the single writer is that it is a serialization point: it cannot be scaled by adding another one.

Autonomy in the reads. Serialization in the writes.