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

Onur Solmaz, founding engineer at OpenClaw and contributor to the Open Claw open-source project, presents how he built acpx — a headless CLI for the Agent Client Protocol (ACP) — and a Kubernetes-based operator for provisioning disposable agent pods, all motivated by the practical challenge of processing 300 to 500 pull requests per day at scale.

Solmaz opens with his background: he has been building AI coding harnesses since before ChatGPT launched, starting with a JupyterLab extension over the original Codex DaVinci Code 2 model. He became a maintainer of the Open Claw project early in its life, has been present through two renamings, and today focuses on agent interoperability, orchestration, and enterprise adoption.

The core problem he describes is that OpenClaw receives over 60,000 total pull requests, with 300–500 new ones arriving daily. Most are AI-generated, often not mergeable, but they all carry signal — user frustration pointing to broken parts of the codebase. Manually triaging them is a classic automatable loop: find intent, judge implementation quality, check for conflicts, run CI, and conduct a review pass. To eliminate himself from this loop, Solmaz built a node-based workflow graph inside acpx that drives a Codex agent session programmatically: reproduce the bug, judge refactor quality, check conflicts, run a review loop, and emit structured JSON output.

ACP (Agent Client Protocol) is contrasted with MCP (Model Context Protocol): where MCP gives tools to a model, ACP standardizes the human-to-agent (and agent-to-agent) communication interface. Solmaz chose ACP because, at the time he needed adapters for both Codex and Claude Code, only ACP had them built. He describes a competing landscape of protocols — including agent-to-agent protocols — and expects eventual multi-protocol support.

acpx is described as a Swiss Army knife CLI for ACP that is slowly accumulating capabilities. It replaces PTY scraping with structured protocol-level communication and allows driving Codex or Claude Code sessions from any client — including Discord channels. Solmaz demonstrates running one to five parallel Codex sessions from Discord channels while traveling, with each channel bound to a distinct agent session via ACP. He built acpx itself through Discord in this way.

The second half of the talk covers a Kubernetes Go operator (hosted at TextCortex's open-source repo, referred to as "sprits") that provisions a full compute environment — one Kubernetes pod per task — wires it into Slack, and tears it down on completion. The motivation is "on-demand disposable agents": rather than a single persistent bot instance per Slack app (which requires manually creating a new app manifest for each new agent identity), the operator dynamically provisions agents with generated names and identities. Solmaz argues that chat platforms (Slack, Teams, Discord) do not yet support multi-agent cosmetic provisioning, which is a blocker for this vision, and that a proper solution requires Kubernetes, an agent harness (Codex, Claude Code, or similar), read/write GitHub access, and a state synchronization mechanism analogous to rsync or Dropbox's sync algorithm.

The operator is currently used at TextCortex for error-reporting workflows: a Slack concierge agent receives a bug report, dispatches a new isolated agent pod to debug it, and the agent works inside a full Kubernetes environment with file edit access. Solmaz draws a favorable comparison to Devin (crediting Cognition for inventing the category) and notes that OpenHands uses Firecracker for VM isolation, while his stack uses full Kubernetes pods — wasteful in resource terms but a better abstraction given that a full compute environment dramatically expands agent capability, as demonstrated by the Open Claw project.

Key architectural components: acpx CLI, ACP adapter layer for Codex and Claude Code, node-based workflow engine, Go Kubernetes operator, Slack/Discord integration, GitHub read/write access for state, and structured JSON output at every workflow node.