Kiro launches an always-on autonomous agent in preview
Kiro has introduced Kiro autonomous agent in preview, positioning it as a persistent developer assistant that maintains context across repositories and tasks. The agent is being rolled out to individual developers on Kiro Pro, Pro+, and Power plans with no charge during the preview period and subject to weekly usage limits. Teams can join the waitlist for early access via the Kiro contact page.
The development context gap
Many existing AI coding assistants are session-based and require repeated context setup when moving between repositories or tasks. That pattern becomes laborious for multi-repo work such as upgrading a widely used dependency. The blog frames the difference across three approaches:
- Manual updates across repositories demand repeating the same steps for each repo.
- Agentic IDE/CLI assistants reduce manual effort but remain session-bound, requiring context re-establishment per repo.
- Kiro autonomous agent treats multi-repo work as a unified task, retaining learned preferences and patterns so architectural choices and code-review feedback apply across subsequent changes.
The key distinction is that the agent is not session-based: it preserves context and learns from code reviews and prior interactions, applying team standards automatically.
How it works
Access to the agent is available through the online Kiro account. Developers can chat to specify changes, then delegate work as tasks. The agent supports up to 10 concurrent tasks and runs each task asynchronously. When a task is assigned, the agent:
- Spins up an isolated sandbox environment mirroring the development setup
- Clones repositories and analyzes the codebase
- Breaks down work and defines requirements and acceptance criteria
- Coordinates specialized sub-agents for research, coding, and verification
- Asks clarifying questions when needed
- Opens pull requests with explanations of changes and decisions
Each task is isolated with configurable network access, environment variables, and development environment settings, allowing the agent to run full test suites and verification steps without blocking human work.
Security, integrations, and environment setup
Sandbox execution includes several configurable controls:
- Network access levels: Integration only (GitHub proxy), Common dependencies (package registries like npm, PyPI, Maven), or Open internet. Custom domain allowlists are supported.
- MCP integrations: Tasks can use Model Context Protocol servers to connect specialized tools and proprietary systems during execution.
- Secrets and environment variables: Secrets are stored encrypted and are not exposed in logs or pull requests.
- Automatic environment configuration: The agent detects DevFiles or Dockerfiles in repositories to set up dependencies and build/runtime commands. If neither is present, the agent analyzes project structure to configure an appropriate environment.
Teams and collective memory
For teams, the agent acts as a shared resource that accumulates a unified memory of codebase standards, discussions, and pull requests. Integration with tools across the stack—GitHub, GitLab, Jira, Slack, Teams, Confluence—enables the agent to incorporate specs and conversational context as work progresses. That shared history allows the agent to apply consistent patterns (naming conventions, error handling, architectural decisions) across parallel tasks and repositories, helping reduce coordination bottlenecks.
Tasks can also be assigned from GitHub by adding the kiro label or mentioning /kiro in issue comments; the agent will monitor issue comments for context and feedback.
Access and next steps
The preview is rolling out gradually to Kiro Pro, Pro+, and Power users. Individual users may sign in to check availability at the agent dashboard, and teams can join the waitlist for early access via the Kiro waitlist page.
Original source: https://kiro.dev/blog/introducing-kiro-autonomous-agent/
