Descriptions:
Onur Solmaz, a founding engineer and OpenClaw maintainer, presents his work building ACPX and deploying open-source agents on Kubernetes at AI Engineer 2026. The talk traces a practitioner’s evolution from running an early Codex-based JupiterLab extension to operating parallel agent workloads across multiple Discord channels—each bound to a separate Codex instance through the Agent Communication Protocol (ACP).
Solmaz distinguishes ACP from MCP: while MCP focuses on giving tools to models, ACP standardizes the interaction layer between agents and client interfaces, allowing editors like Zed, VS Code, and others to support any agent through a single integration rather than building redundant plugins. ACPX, his personal project, binds Discord channels directly to Codex via ACP, enabling mobile-friendly coding workflows—including a live demo of converting ACP documentation to PDF from an airport lounge before a flight.
A substantial portion of the talk addresses a real challenge in large open-source projects: absorbing thousands of community pull requests, many AI-generated and poorly scoped. Solmaz describes building abstract agent workflows that automatically classify PR intent, judge implementation quality, check for merge conflicts, enforce CI, and loop on superficial fixes before human review. The session provides an honest, unpolished view of what agent orchestration looks like when used to manage a project with tens of thousands of stakeholders—including the tradeoffs between autonomous fixing and knowing when to escalate to a human.
📺 Source: AI Engineer · Published May 21, 2026
🏷️ Format: Workflow Case Study
