Skip to content

SYSTEM Cited by 1 source

Kiro-flock

Kiro-flock is AWS's open-source reference implementation of a self-organizing multi-agent cluster. Independent headless Kiro CLI sessions run on separate EC2 instances and coordinate by reading and writing an S3 shared environment; there is no work scheduler, message bus, or central result aggregator. (Source: sources/2026-08-11-aws-scaling-patterns-for-self-organizing-multi-agent-clusters-with-kiro)

Architecture

The shared environment contains one direction file, one append-only log per agent, and a shared artifact area. In each iteration, an agent starts with a fresh session, reads the direction plus a configured peer set, selects a contribution, writes its artifact, and appends a structured action / result / next_intent record to its own log.

A management plane behind API Gateway, Lambda, and Cognito starts, stops, and steers clusters. CloudWatch collects metrics, and Bedrock supports post-run analysis. The management plane does not make task assignments. (Source: sources/2026-08-11-aws-scaling-patterns-for-self-organizing-multi-agent-clusters-with-kiro)

Coordination algorithms

  • Amorphous — fixed-radius logical ring; constant peer-read work as the cluster grows, intended for diverse parallel exploration.
  • Mesh — every agent reads every peer's latest entry; rapid convergence for smaller groups at the cost of diversity and linearly growing context.
  • Swarm — each agent reads the most recently active peers; tracks active work, but can hot-spot if the recency set is too small.

The reference implementation supports runtime topology changes, enabling the amorphous → swarm → mesh sequence. AWS reports a largest run of 184 agents across 11 cooperating clusters; the implementation is explicitly a sample rather than a production service. (Source: sources/2026-08-11-aws-scaling-patterns-for-self-organizing-multi-agent-clusters-with-kiro)

Cluster composition

The source's WeltenBuilder example applies the same shared-environment rule above the individual-agent level: feature, shared-infrastructure, QA, and coordinator clusters publish artifacts that peers read on later iterations. A coordinator's conflict-resolution note is a signal rather than a command, so removing it reduces guidance without creating a hard dependency. See patterns/cluster-of-clusters-artifact-coordination.

Last updated · 619 distilled / 1,953 read