Silen Naihin is the co-founder and chief technology officer of Experiential Labs, which helps companies turn everyday AI usage into specialized models they own. Previously a founding engineer at AutoGPT, he builds infrastructure for evaluating autonomous agents, supervising their actions, and improving them through experience.
Naihin studied computer science at Minerva University before leaving his degree program. At AutoGPT, he helped build an open-source agent project that reached 160,000 GitHub stars and led work on Auto-GPT-Benchmarks, which measured agent performance across implementations. He subsequently co-founded Stackwise, a Y Combinator Winter 2024 startup, and worked with the U.S. Department of Energy on applying AI to scientific discovery.
In 2026, Naihin co-founded Experiential Labs with Kion Fallah; the company joined Y Combinator’s Summer 2026 batch. Its open-source model gateway connects proprietary, open-source, local, and custom models. Production traces and simulated tasks help train specialized alternatives, while difficult requests can still go to frontier systems. Announcing the launch, Naihin argued that rising AI expenditure should create assets companies own.
- Agent evaluation as continuous feedback. Naihin developed benchmarks and Agent Protocol-compatible tooling to assess whether software changes actually improved autonomous agents. He treats noisy measurements as useful engineering signals, without confusing them with definitive scientific findings.
- Action-level agent safety. As first author of 2023 research introducing AgentMonitor, Naihin proposed auditing agent behavior, interrupting unsafe actions, and escalating suspicious activity for human review. His concerns include prompt injection through malicious webpages and destructive mistakes by agents with excessive system permissions.
- Continual learning from production experience. In CLaaS research co-authored with Fallah, Barak Widawsky, and Qingqing Mao, Naihin examined asynchronous training and experience replay to improve deployed agents, transfer knowledge between tasks, and reduce forgetting.
- More interpretable model features. With Lev Stambler, Naihin developed cosine-scored sparse autoencoders, addressing how conventional scoring can mistake activation magnitude for meaningful features by combining directional similarity with learned magnitude dependence.
