What runner did Amazon open?
Amazon has released Kiro Crew, the system it built internally as MeshClaw to run several coding agents at once, under the Apache 2.0 licence. It runs on macOS, Linux and Windows and connects through the Model Context Protocol, webhooks and chat tools including Slack. I think the break sits in the runner that keeps agents working across sessions.[1]
Kiro Crew keeps agents working across sessions with shared memory and reusable skills, and divides work between them through the Agent Client Protocol. That layer ties how one request moves from model to tool and from one agent's output to the next agent to a protocol. The honest previous baseline was a single-agent flow that restarted each session; shared state and protocol-based division are now visible.[1]
What cost arrives with parallel agents?
On the security side the listed controls include operating-system sandboxing, a deny-by-default command list, blocking of suspicious patterns, input validation and audit logging. Those controls make auditability part of the design while agents run in parallel; they do not by themselves erase the conflict, retry and supervision load that parallel runs bring. The open-source licence makes the code inspectable; reading logs and rejecting commands still lands on the operator.[1]
Amazon says more than 39,000 of its own developers and 500 contributors used the internal version within six months; the adoption figures come from Amazon's own statement and no independent count is reported. For me those numbers are an in-house adoption claim, not proof of habit. For an outside team the baseline is still zero: pulling MeshClaw as Kiro Crew and running it in your own repository is possible, but seeing the same coordination payoff in production takes a separate measurement.[1]
What should builders watch outside Amazon?
The observable signal is whether teams outside Amazon move this runner into production workflows. If independent repositories document multi-agent runs through the Agent Client Protocol and shared skills persist across sessions, value is accumulating in the orchestration layer rather than in model choice alone; otherwise the open-source release stays a showcase that does not mirror internal MeshClaw scale. By 31 December 2026 a count of public production examples built on Kiro Crew makes that distinction measurable.[1]