AWS open-sources Kiro Crew to run AI agents in the background

Amazon announced Kiro Crew, an open source system for running multiple Kiro code agents asynchronously, across sessions, tools, and tasks. The idea is to move away from the one-prompt-at-a-time model and delegate work that keeps running without active supervision: incident investigation, ticket triage, migrations, and pull request monitoring.

The project was released under the Apache 2.0 license and is available for macOS, Linux, and Windows, with integrations for Slack, Telegram, and WeCom. It runs on top of the Kiro CLI and reuses existing .kiro configurations (steering files, skills, and custom agents), which lowers the barrier to entry for those already using the Kiro ecosystem.

From internal tool to public project

Kiro Crew wasn't born as a product. According to Amazon, it was developed internally under the name MeshClaw and reportedly adopted by more than 39,000 developers within the company before becoming open source. The authors Bolin Chen (senior software engineer), Zejiang Guo (principal SDE at AWS), and Zezhen Xu (senior software engineer) describe the motivation in fairly pragmatic terms:

The three of us wanted something simple that wasn't available internally: a way to kick off a task, walk away, come back to something worth reviewing, and run several tasks at once instead of babysitting one prompt at a time.

-- Bolin Chen, Zejiang Guo, and Zezhen Xu, Amazon engineers

They cite inspiration from OpenClaw and tools with self-learning agents, but stress the need to meet security requirements for internal development work. Organic adoption was the point most highlighted by the community. Muhammad Ishaq, CTO of Nextbridge, commented:

The internal adoption numbers say more than the feature list, honestly. 39,000 builders and 500 contributors in six months without a mandate means people were actually solving real problems with the tool, not just testing something new because it existed.

-- Muhammad Ishaq, CTO of Nextbridge

How it works in practice

Kiro Crew supports persistent, multi-session work through a few pillars:

• Shared memory, so agents retain project context between sessions;
• Reusable skills and scheduled jobs;
• Concurrent agents, capable of delegating work to subagents;
• Purpose-built apps;
• Integration with external tools via MCP (Model Context Protocol) and webhooks.

Agent orchestration happens through the Agent Client Protocol (ACP), which gives live visibility into what each agent is doing. The so-called Activity view shows each agent's plan, tool calls, approval gates, and results as they happen, which matters when you have several tasks running in parallel without someone watching all the time.

An important detail for teams concerned with standardization: Kiro Crew supports skills built for other open, standards-based agent platforms without modification. In other words, the bet is on interoperability on top of protocols like MCP and ACP, rather than a closed ecosystem.

Security as a central argument

Security appears as an explicit differentiator, and not by chance: agents that execute commands and touch code without supervision are a risk surface. The authors list a defense-in-depth approach:

Kiro Crew brings defense in depth from day one: OS-level sandboxing, deny-by-default commands, blocking of suspicious patterns, input validation, blocking of sensitive paths, credential redaction, and a signed audit log of every action.

-- Bolin Chen, Zejiang Guo, and Zezhen Xu, Amazon engineers

This set (OS sandbox, denied-by-default command policy, credential redaction, and signed audit log) is the kind of control usually required to put AI automation near production repositories and pipelines.

The point of concern: token consumption

Not everything is enthusiasm. While many developers report experiments with CI/CD migrations, Dependabot-based triage, scheduled tasks, and long-running work, others have raised cost concerns, pointing out that Crew can burn through tokens significantly faster than the Kiro CLI. That makes sense: running multiple concurrent agents, accumulating context across sessions, tends to multiply model calls.

On the positive side of that same mechanic, Mathi M, senior solution architect at Verizon, assesses:

Having background subagents handle parallel jobs without blocking the main flow is exactly what builders need. Stacking context across sessions instead of starting from scratch will save a lot of redundant prompt engineering.

-- Mathi M, senior solution architect at Verizon

It's a classic trade-off: you save engineering time and prompt rework, but you pay in tokens. For Brazilian teams operating with an unfavorable exchange rate, that token consumption bill isn't a minor detail, it's a decision factor. The fact that Kiro Crew can run locally or on infrastructure controlled by the team itself helps on the governance and privacy side, but it doesn't insulate the cost of the model behind the agents.

What this changes for developers in Brazil

For Brazilian developers, Kiro Crew fits a trend that had already been showing up with tools like Cursor and agent-native alternatives: moving AI work from a synchronous assistant (you ask, it answers) to agents that run in the background and hand back something reviewable. That changes the workflow design, especially for repetitive maintenance tasks (ticket triage, dependency updates, large migrations) that today compete with features in the backlog.

In practice, the safest evaluation path would be to start with low-risk, well-scoped tasks, backed by the approval gates and the signed audit log, before granting autonomy to agents in critical repositories. And to closely monitor token consumption from the pilot stage on, given the community's own warning.

Some points remain open. Amazon says Kiro and AWS engineers will maintain the project initially, with external maintainers potentially joining as contributors emerge, and that it expects community participation in development and roadmap planning. How this open source governance model will hold up (and how much of the product depends on paid AWS services behind it) is what the coming months should clarify.