Onur Solmaz

Onur Solmaz is an engineer at Hugging Face, an OpenClaw maintainer, and the creator of acpx, a shared interface for coordinating coding agents across tools and models. He builds infrastructure that makes agents interoperable, deployable on systems their users control, and capable of handling substantial engineering workflows under human supervision.

Solmaz studied civil engineering at Middle East Technical University and earned a master’s degree in computational mechanics from the University of Stuttgart. In 2022, before ChatGPT launched, he developed icortex, an experimental coding agent built around an early Codex model and a custom Jupyter kernel. The project evolved into TextCortex’s AI harness and main chat product.

A founding engineer at TextCortex, Solmaz began as its sole backend developer and later became vice president of engineering and research. He introduced Kubernetes-based development infrastructure, isolated branch previews, elastic CI runners, and short-lived cloud credentials; his account of the overhaul reports an 80 percent reduction in CI costs.

After adopting OpenClaw shortly after its initial Discord release, Solmaz contributed enterprise integrations, including Microsoft Teams, and became a maintainer. His responsibilities expanded into subagents, security, browser and terminal use, and coding-agent interoperability. At Hugging Face, he focuses on making OpenClaw and other agent harnesses effective with local and open models.

• One protocol, many coding agents. Solmaz created acpx, a headless client for the Agent Client Protocol, supporting persistent sessions, permissions, structured output, and parallel workflows. By connecting Discord channels directly to separate Codex sessions, he avoids intermediaries paraphrasing instructions and demonstrates how one interface can serve multiple agents.

• Automated review with human architectural oversight. His agent workflows infer pull-request intent, reproduce bugs, inspect feedback, resolve straightforward conflicts, and escalate fundamental design decisions. With pi-workflows, he extends this into graph-based execution; his reactive workflow experiments add planning, explicit approval requests, and priority-aware review.

• Disposable agents with isolated infrastructure. At TextCortex, Solmaz developed Spritz, a Kubernetes-native control plane that assigns tasks dedicated containerized agents. Its Slack-triggered debugging workflows expose a practical limitation: enterprise chat platforms still make independently identifiable, dynamically provisioned agents difficult to deploy.

• Open models need standardized evaluation. His projects include ML Claw, which deploys OpenClaw agents on Hugging Face infrastructure, and localperf, which benchmarks local inference. He argues for standardized agent evaluation under common harness conditions so model performance can be distinguished from proprietary orchestration advantages.