Inkitt Develops Resumable Agent Orchestrator for Automated Pull Requests
New event-driven architecture automates low-risk code changes while maintaining strict security boundaries.
1Automated Pull Request Workflow
Inkitt has implemented a one-shot pull request creator designed to automate minor, low-risk software modifications. The system handles the entire lifecycle from initial task intake to generating a draft pull request ready for human review. By operating on remote infrastructure, it eliminates the need for developers to manage coding agents locally.
2Event-Driven Agent Orchestration
Lead Engineer Tanner Letkiewicz describes the system as an event-driven orchestrator that manages short-lived agent containers. Unlike rigid pipelines, this architecture coordinates individual stages including specification, implementation, and verification. The orchestrator persists every execution state in a database, allowing the workflow to resume seamlessly from its last recorded point if interrupted.
3Separation of Concerns Architecture
The orchestrator maintains a strict separation between decision-making logic and external side effects. It maps current run states to next actions without direct database or network access. Agent containers receive only short-lived repository tokens, ensuring they cannot access sensitive cloud identities, Slack credentials, or arbitrary network connections during the automated coding process.
4Security and Scope Constraints
Automation is restricted to predefined boundaries to mitigate risks. The system explicitly rejects complex tasks such as database migrations, schema changes, authentication updates, or package modifications. A dedicated security check inspects every Git diff to ensure that changes do not touch restricted file paths, maintaining integrity across the entire codebase during execution.
5Idempotency and Workflow Reliability
To handle retries and potential duplicate webhook deliveries, the system applies idempotent operations. It utilizes conditional database writes to prevent errors from repeated events. For CI completion, the orchestrator queries GitHub directly to verify that all relevant jobs have finished, rather than relying on individual job events to signal overall workflow completion.
6Comparison with Existing Models
While sharing safeguards with tools like GitHub Copilot, Inkitt’s model differs through its focus on persistent workflow state and centralized control. Unlike SWE-agent, which utilizes a model-driven interface for browsing and editing, Inkitt prioritizes bounded retries and specific orchestration patterns to manage external actions securely within a controlled, remote development environment.
7Mandatory Human Approval Process
Despite the high level of automation, human oversight remains a critical component of the workflow. The system cannot bypass branch protection or merge its own changes. Once a draft PR is created and CI passes, a human must manually mark the PR as ready for review and manage the final approval process.
8Operational Impact and Future
By shipping code daily through this automated system, Inkitt demonstrates a shift toward more efficient, agent-assisted development cycles. The architecture provides a scalable framework for handling repetitive tasks while keeping developers focused on complex engineering challenges. This approach balances the speed of AI-driven automation with the necessary safety of human-led code reviews.
Original article: InfoQ — Inkitt Develops Resumable Agent Orchestrator for Automated Pull Requests