PATTERN Cited by 1 source

Shared-state agent coordination¶

Pattern¶

Give independent agents one durable shared environment containing a goal, append-only contribution records, and artifacts. Each agent reads a selected subset of the environment, decides its next contribution independently, writes its output, and records an advisory result for peers. No component owns task assignment, aggregation, or completion. (Source: sources/2026-08-11-aws-scaling-patterns-for-self-organizing-multi-agent-clusters-with-kiro)

agent → read goal + selected peer traces → choose work → write artifact + append trace
  ↑                                                                    │
  └──────────────────── later peers observe and react ────────────────┘

Use when¶

The work divides into many quasi-independent contributions.
Alternative approaches are valuable and some duplication is acceptable.
Participants may join or fail over a long-running run.
A final validation gate is sufficient; no gate is needed between each contribution.

Do not use when¶

The task tree and dependency order are known in advance.
A central reviewer must approve each intermediate action.
Low interactive latency or strict sequencing is the dominant concern.

Design constraints¶

Keep state explicit and durable; participants should not depend on a previous agent's process memory.
Bound what each agent reads to manage context and preserve diversity; see concepts/bounded-peer-visibility.
Archive or namespace old environments so stale artifacts do not become current signals.
Add an external test, review, or policy gate for outcomes that cannot tolerate unverified intermediate work.

Contrast: broker-mediated coordination¶

patterns/multi-agent-streaming-coordination uses a broker for ordered consumer semantics, durable fan-in/fan-out, and explicit event delivery. This pattern uses shared state as a read-and-react substrate instead: an agent can see a trace without being assigned it, and no broker or aggregator is required. The trade-off is weaker workflow control.

Seen in¶

systems/kiro-flock — Kiro CLI agents on EC2 coordinate through an S3 direction file, per-agent logs, and shared artifacts.