Source: AWS Architecture Blog — 2026-08-12
Summary
AWS published a reference architecture, kiro-flock, for running clusters of Kiro CLI agents on EC2 that coordinate entirely through an S3 bucket instead of a central orchestrator or message bus. Each agent is a headless Kiro session on its own EC2 instance, reading and writing append-only logs in shared S3 state to converge on work without any process directing them. It's a direct alternative to the hub-and-spoke orchestrator pattern most multiagent frameworks default to.
Key Takeaways
- There is no orchestrator, no message queue, and no central coordinator process in this architecture — agents self-organize purely by reading each other's append-only logs in a shared S3 bucket.
- Each agent runs as an independent headless Kiro CLI session on its own EC2 instance, meaning the cluster scales by adding instances, not by scaling a central bottleneck.
- The append-only log pattern in S3 gives the cluster an implicit audit trail and lets new agents join mid-run by simply reading the accumulated shared state.
- This is a genuinely different scaling pattern from the request-routing/orchestrator model most agent frameworks assume by default — worth evaluating when a central coordinator becomes the bottleneck or single point of failure in your own multiagent system.
Reel Script
Hook: Most multiagent systems have a traffic cop — one orchestrator process routing every task. AWS just published an architecture that deletes the traffic cop entirely and replaces it with an S3 bucket.
Core Concept:
The standard way to run multiple AI agents together is a hub-and-spoke model: a central orchestrator process assigns tasks, tracks state, and routes results. It works, but that orchestrator becomes a single point of failure and a scaling bottleneck — every agent has to talk to it. AWS's kiro-flock reference architecture takes a completely different approach borrowed from distributed systems design: shared-state coordination. Instead of agents talking to a central brain, every agent — literally just a headless Kiro CLI process running on its own EC2 instance — reads and writes to a shared S3 bucket using an append-only log. Think of it like a group project where instead of a project manager assigning tasks, everyone just checks a shared whiteboard, sees what's already claimed, and picks up what's left undone.
Hands-On: Here's the actual coordination mechanic: each agent, on its own EC2 instance, periodically reads the S3 bucket to see what other agents have logged — completed work, claimed work, discovered facts. It then decides its own next action based on that shared state and appends its own log entry back to S3. No agent ever calls another agent directly. No central process assigns work. New agents can join a running cluster mid-task just by reading the accumulated log and inferring what's left to do. The append-only property means you get a full audit trail for free — you can replay exactly how the cluster converged on its final state just by reading the log in order, which is the kind of debuggability that a live orchestrator with in-memory state simply doesn't give you.
Takeaway: If you've hit a wall with your orchestrator becoming a bottleneck as your agent fleet grows, this pattern is worth stealing even outside AWS — the core idea (shared append-only state instead of a central router) works with any object store. It trades a bit of coordination latency for horizontal scalability and a built-in audit log, and for a lot of "swarm of agents working a big batch job" use cases, that's the right trade. Go read the reference implementation before you build your next orchestrator.