Anyone who's used Cursor or Copilot to write code has probably had this moment of frustration: close the browser tab, quit the IDE, and that agent you spent all afternoon pairing with is just gone. Open it up tomorrow, and you've got a brand-new AI that remembers nothing.

You're the integration layer. The moment you leave, the process stops.

On September 7, Amazon Web Services open-sourced Kiro Crew — a multi-agent collaborative system built to fix exactly this pain point. Its core mission is simple: let coding tasks run persistently across sessions. Drop the task, walk away, come back to results. Born from Amazon's internal MeshClaw project, it already serves over 39,000 developers internally.

From Vibe Coding to Async Hosting — What's the Difference?

Most AI coding tools today are still stuck in the "vibe coding" phase: you sit in front of the screen, chat with the agent line by line, it writes a bit, you review a bit. Your productivity goes up, but your time is still tied to the process.

What Kiro Crew does is turn coding agents from "chat window assistants" into "background workers." It supports cross-session state preservation — meaning you can drop a refactoring task at 5pm, shut down your computer, and see results the next morning. In between, it plans its own steps, calls its own tools, retries on its own when things fail, and keeps making progress.

Three key design choices make this work:

First, shared memory and reusable Skills. Multiple agents can share the same project context and knowledge base — no need to re-feed the codebase every time. Skills can be encapsulated and reused, just like function calls.

Second, multi-agent parallel scheduling. A big task can be split into subtasks, dispatched to several agents working simultaneously — one writing tests, one updating docs, one doing refactoring. Results get merged at the end.

Third, scheduled triggers and Webhook integration. Agents can run automatically at specific times (like a nightly code audit), or connect to external systems via MCP and Webhooks — auto-investigate when CI breaks, auto-triage when a new issue is filed.

The Engineering Path Validated by 39,000 Internal Users

Kiro Crew isn't a lab demo — it's a production-grade tool that's been running inside Amazon for a while.

39,000 internal developers means it's been battle-tested at real scale — the size of the codebases, the diversity of tasks, all the weird ways things can go wrong — none of that can be simulated by a small team experiment. Darko Mesaroš, AWS Principal Developer Advocate, put it this way: long-running doesn't mean letting the model loop forever. It means tasks can progress across multiple sessions while preserving necessary context, execution logs, and human intervention points.

The most important detail here is "human intervention points." A lot of people hear "async agent" and panic about loss of control — what if the agent runs amok? What if it deletes code? Kiro Crew's answer is fine-grained security control and observability: every operation is logged, critical junctures require human approval, and you can step in anytime something goes wrong.

That's what engineering-grade AI coding tools look like. Not flashy demos — but manageable, controllable, auditable productivity tools that actually land in real teams.

The first act of AI coding was vibe coding — you sit at the screen, pair with an agent, efficiency improves, but you're still the bottleneck. The second act is starting now: async hosted agents running in the background, humans moving from supervisors to decision-makers, finally being freed from repetitive work.

Kiro Crew going open source is just the starting point. As more agents step out of chat boxes and into persistent background execution, the way software development gets done will be redefined.

See you tomorrow.