Scaling Autonomous Agent Workflows on Kubernetes

Automating PR Triage with acpx

OpenClaw receives 300 to 500 pull requests daily, most of which are AI-generated and non-mergeable. To handle this volume without manual intervention, Onur Solmaz developed acpx, a headless CLI tool that implements the Agent Client Protocol (ACP). Instead of relying on PTY scraping, acpx uses structured communication to drive agent sessions through a node-based workflow graph. This workflow automates the entire lifecycle of a PR: reproducing the reported bug, judging the implementation, checking for merge conflicts, and running iterative review loops. The system emits structured JSON output, allowing developers to integrate these agent-driven insights into broader CI/CD pipelines.

Kubernetes-Based Agent Orchestration

To move beyond the limitations of single-instance chat integrations like Slack or Discord, Solmaz designed a Kubernetes operator that provisions disposable compute environments for individual tasks. This architecture treats each agent as a full-compute pod, providing the agent with a dedicated environment to read, write, and synchronize state. By decoupling the agent from specific chat platforms, the system allows for multi-agent provisioning where each task runs in its own isolated container. When a task is completed, the operator tears down the pod, ensuring efficient resource usage while maintaining the power of a full-machine environment for the agent.

Standardizing Agent-to-Client Interaction

The Agent Client Protocol (ACP) serves as the foundation for this interoperability. Unlike the Agent Protocol, which focuses on agent-to-agent communication, ACP standardizes how humans and clients interact with agents. By adopting this standard, developers avoid the redundant work of building custom plugins for different IDEs or chat interfaces. Solmaz emphasizes that as these protocols mature, the industry will move toward a model where agents are applied generously to any task that can be solved programmatically, effectively removing the human from the loop for mechanical coding and triage tasks.