Scaling Agents on Kubernetes with acpx and ACP — Onur Solmaz, OpenClaw

Key Concepts

• ACP (Agent Client Protocol): A standard for agent-to-client interaction, designed to unify how humans interface with various AI agents.
• Open-Source Agent Frameworks: Tools like Open-Claude (Open Claw) that allow developers to build and deploy autonomous coding agents.
• Kubernetes-based Orchestration: Using container orchestration to manage the lifecycle, scaling, and state of AI agents.
• Agent Interoperability: The ability for different agents (e.g., Claude, Codex) to communicate and work together through standardized protocols.
• TDD (Telegram/Discord Driven Development): A methodology where development workflows are managed and executed via chat platforms.
• Standard Operating Procedures (SOPs) for Agents: Automated, structured workflows that define how agents should handle tasks like PR reviews, bug reproduction, and refactoring.

---

1. Main Topics and Key Points

The presentation focuses on building scalable, interoperable AI agent systems for enterprise environments. The speaker highlights the transition from manual, fragmented agent usage to a standardized, automated infrastructure.

• The Challenge of Scale: Open-source projects like Open-Claude face massive influxes of PRs (300–500 per day). The core problem is how to process this volume without creating "AI slop" (low-quality, automated noise).
• Standardization via ACP: The speaker advocates for ACP over competing standards because it provides necessary adapters for tools like Claude Code and Codex, reducing duplicated effort across different IDE plugins.
• Automating the Automator: The speaker proposes a framework where agents are not just used for coding, but for managing the process of coding—judging PRs, resolving conflicts, and running CI/CD loops.

2. Important Examples and Applications

• Discord/Telegram Driven Development: The speaker uses Discord channels as a "full IDE," running parallel workloads where multiple agents (Codex) handle different tasks simultaneously.
• ACP-X: A "Swiss Army knife" tool built by the speaker to allow agents to call other agents via the command line, facilitating complex, multi-step workflows.
• Error Reporting Workflow: A real-world application where an agent is dispatched via Slack to debug production issues, which then spins up a dedicated Kubernetes pod to handle the task.

3. Step-by-Step Methodologies

The speaker outlines a framework for managing PRs using AI:

1. Intent Identification: Analyze the PR description to determine the user's goal.
2. Implementation Judgment: Evaluate the code quality and check for conflicts.
3. Automated Review Loop: If the CI fails or the code is suboptimal, the agent performs a "refactor loop."
4. Human Escalation: If the agent identifies a fundamental design issue that requires human intuition, it flags the PR for manual review.
5. Structured Data Output: All agent feedback is converted into JSON-structured data to ensure it can be processed by other systems.

4. Key Arguments and Perspectives

• Agents as "Ointment": The speaker argues that agents should be applied "generously" to any problem that can be solved through automation, effectively taking the human out of the mechanical loop.
• Enterprise vs. Personal Agents: The speaker predicts a divergence where enterprise agents will consume significantly more inference and require robust, on-demand, disposable infrastructure (Kubernetes) compared to personal agents.
• The Need for Multi-Agent Provisioning: Current chat platforms (Slack/Discord) lack the ability to easily spawn and manage multiple distinct agent identities. The speaker argues that platforms must evolve to support "cosmetic" multi-agent provisioning.

5. Notable Quotes

• "I’m a believer that... running an agent in a loop doesn't necessarily have to be something that will create slop. As long as you’re not making it design something, but you’re making it uncover shallow bugs... it should be fine."
• "You need to take yourself out of the loop and solve it with agents."

6. Technical Terms and Concepts

• Codex: An AI model used as a primary "harness" for coding tasks.
• Helm Charts: Used for deploying the agent-based applications into Kubernetes clusters.
• Firecracker: A virtualization technology (mentioned in the context of OpenAI) used for secure, lightweight isolation of agent environments.
• Rsync/State Sync: Methods used to ensure that files created or edited by agents across different environments remain synchronized.

7. Synthesis and Conclusion

The presentation concludes that the future of AI-assisted development lies in standardized interoperability and infrastructure-level orchestration. By moving away from manual, fragmented agent interactions and toward a Kubernetes-backed, protocol-driven (ACP) architecture, developers can handle the massive scale of modern open-source contributions. The speaker emphasizes that while the technology is still in its "work in progress" phase, the shift toward disposable, on-demand agents is inevitable for enterprise efficiency.