Parth Sareen is a software engineer at Ollama, where he leads work on agent tooling and created Ollama Launch. His work makes open models more useful in everyday applications by improving the software that connects them to tools, local computers, and real-world tasks.
Sareen studied mechatronics engineering at the University of Waterloo before working on Siri, on-device machine learning, and speaker diarization at Apple, and on distributed systems for energy products at Tesla. He began experimenting with language-model agents in 2023 and co-founded Extensible AI with Omkaar Kamath in early 2024. The startup joined Betaworks’ accelerator for AI-agent companies and developed open-source logging and observability infrastructure to improve agent reliability.
Its projects included DAGent, a framework that organizes agent behavior using directed acyclic graphs. As models became more capable of selecting their own actions, Sareen shifted his attention toward designing the surrounding environments that determine whether those actions succeed.
Agent harnesses shape model behavior. Tool definitions, execution loops, permissions, conversation history, context compaction, and caching influence what an agent can accomplish. Sareen argues that open models can appear less capable when their harnesses assume the larger context budgets and behavior of proprietary frontier systems.
Agent reliability requires usable context. Sareen favors straightforward execution loops, clearly defined tools, bounded outputs, and careful management of older context. His approach to reliable agents prioritizes evaluation against the actual tasks an application must complete.
Ollama Launch removes integration friction. Sareen built the product to connect open models with coding and personal agents while handling configuration, model switching, and differing application requirements behind a simpler interface.
Model quantization must preserve practical usefulness. On the Compression at the Edge panel, Sareen described testing compressed models inside agent harnesses before selecting default precision levels. Smaller models sometimes need higher precision to remain effective, and emerging architectures require separate implementation and validation. He also anticipates greater adoption of quantization-aware training by model developers.
Sareen also maintains watchy, an agent-assisted background-task manager that can inspect logs and manage processes, and ducky, a terminal assistant that converts natural-language requests into shell commands.
Making a model fit is only the beginning: selective precision, architecture-aware evaluation and runtime compression determine whether local intelligence remains useful.
