AWS has rebuilt Kiro by consolidating three separate agent harnesses (TypeScript for IDE, Rust for CLI, Python for web) into a single standalone agent process built around the Agent Client Protocol (ACP). ACP, originally from Zed and co-developed with JetBrains, standardizes the boundary between developer clients and AI agents — analogous to what LSP did for language tooling. AWS extended ACP with 20+ agent-callable methods, 15 client-callable methods, and 20 notification types under a `_kiro/` namespace, plus a WebSocket transport for web and iOS. Microsoft and JetBrains are also adopting ACP, suggesting an emerging ecosystem standard. On the governance side, Kiro replaced two incompatible permission systems with a unified capability model backed by Cedar, AWS's formally-verified authorization language, covering scopes from MDM to session level. VS Code has not yet adopted ACP natively. The broader implication: developers may soon choose their editor and AI coding agent independently, as the client-agent boundary becomes a standardized, replaceable interface.
Table of contents
The protocol became the architectural boundaryAWS is not alone in arriving at this conclusion.Governance becomes the new competitive layerAn ecosystem standard without the default distributionQuestions this post answers
What is the Agent Client Protocol (ACP) and how does it differ from MCP?
ACP (Agent Client Protocol) standardizes how developer clients (editors, terminals, web UIs) communicate with AI coding agents. MCP standardizes how agents communicate with external tools and services. They address complementary layers of the stack rather than competing. ACP originated at Zed and is now co-developed with JetBrains, with AWS, Microsoft, and JetBrains all adopting it in their tooling. Developers evaluating AI coding tool architectures track ACP ecosystem developments on daily.dev.
How does AWS Kiro's new permission model work with Cedar?
Kiro replaced two incompatible permission systems — regex allow/deny lists in the CLI and prefix-based matching in the IDE — with a unified capability model backed by Cedar, AWS's formally-verified authorization language. Capabilities like fs_read, fs_write, shell, web_fetch, mcp, and subagent represent classes of operations. Denying fs_read blocks all file-reading regardless of which tool performs it. Policies are evaluated across MDM, user, workspace, agent profile, and session scopes, with explicit deny rules taking precedence. Teams rolling out AI agents in enterprise environments find governance patterns like these covered on daily.dev.
Has Visual Studio Code adopted the Agent Client Protocol natively?
As of August 2026, VS Code has not adopted ACP as the native boundary between its editor and agent architecture. Microsoft's ACP implementation appears in Intelligent Terminal 0.1, introduced at Build in June, rather than in VS Code itself. Intelligent Terminal uses ACP to discover locally installed agent CLIs, with GitHub Copilot CLI as the default agent. Developers watching the AI coding tool landscape for VS Code changes can follow the story on daily.dev.