One Agent for Every Environment: How Kiro Built Its Agent Harness

Kiro unified the agent harnesses of the IDE, CLI, and web to provide a single, continuous agent session across every environment developers use — laptop, cloud sandbox, mobile, and Slack.

Previously, each client ran a separate harness. The IDE was implemented in TypeScript, the CLI in Rust, and the web in Python, each with different session storage methods, tool sets, and configuration models.

As a result, permission systems and command-matching approaches were also incompatible. The CLI used regex-based allowedCommands/deniedCommands, while the IDE applied prefix matching for trustedCommands along with separate substring matching for its denylist. Compaction, sub-agent context sharing, and custom agent behavior also differed by client.

Features were fragmented as well.

• Spec-driven development and powers were available only in the IDE
• Plan mode and code intelligence were available only in the CLI
• New features and bug fixes had to be repeated across three codebases

The team also considered simply aligning the behavioral contract between clients, but every new feature would then require writing a spec, building three implementations, and verifying that behavior matched across all of them. Ahead of the web launch, Kiro decided to build a single agent harness to reduce this duplicated cost and deliver identical functionality to every client.

Rather than compiling into each client as a library, the new harness was designed as a standalone server process. This reduces the risk of clients depending directly on internal methods or bolting on their own logic, and establishes a foundation for maintaining consistent agent behavior across different surfaces and execution environments.