Vincent Koc is Chief Architect of the OpenClaw Foundation and a core maintainer of OpenClaw, the open-source personal AI agent. He designs the architecture, evaluation, observability, and security systems that make autonomous agents easier to supervise as they acquire greater independence.
Koc previously held data-leadership positions in Australia: digital analytics manager and analytics chapter lead at Qantas in 2020, followed by head of data at hipages Group by 2023. He subsequently worked at Comet on AI research, evaluation, and developer relations, investigating how non-Latin languages expose weaknesses in conventional model benchmarks and introducing Opik Agent Optimizer in 2025.
Dark-factory agent orchestration. Koc organizes coding agents into parallel lanes for bug fixes, features, continuous integration, and urgent issues, with engineers supervising priorities and reasoning quality. His OpenClaw maintenance workflow treats human judgment as the limiting resource, not raw commit volume.
Modular agent architecture. While collaborating on NVIDIA NemoClaw, Koc helped refactor OpenClaw around plugins and separable provider integrations, reducing pressure to absorb every community feature into the core codebase.
Malleable evaluations. He argues that agents with changing tools, users, and objectives cannot be adequately measured by fixed benchmarks. His adaptive approach to agent evaluation combines component testing, production traces, refreshed evaluation suites, and feedback aligned with user intent.
Telemetry-aware agent optimization. Koc connects operational evidence directly to corrective action: harnesses can use failures, costs, and evaluation results to adjust behavior. His tokenjuice project extends that concern to token efficiency in terminal-heavy coding workflows.
Layered security for agent skills. As lead author of ClawHub Security Signals, Koc examined 67,453 public OpenClaw skill versions using complementary security checks. The resulting silver-standard security dataset reflects automated judgments, not human-verified maliciousness.
When agents generate changes faster than maintainers can read them, the engineering work shifts toward task allocation, regression signals, reusable skills, and human judgment.
As agents adapt their tools, behavior and user experience, evaluation needs to evolve with them—from fixed examples to trace-derived tests and telemetry-guided feedback.
▶ Watch ↗