Ishaan Sehgal

Ishaan Sehgal is the co-founder and chief executive of Omnara, an open-source control plane for AI agents. He builds infrastructure that keeps agents operational when machines fail, permissions require human intervention, or users switch models and providers.

Sehgal earned bachelor’s and master’s degrees in computer science at the University of Illinois Urbana-Champaign, interned at Roblox, and worked on machine learning at Windsor before joining Microsoft’s Azure Kubernetes Service organization. At Microsoft, he became a core contributor to KAITO, the Kubernetes AI Toolchain Operator, which helps developers deploy and fine-tune models as cloud-native workloads. His work on memory-efficient fine-tuning explored quantization, QLoRA, batch sizing, and PyTorch allocation to make adapting Phi-3 models practical on constrained hardware.

In 2025, Sehgal co-founded Omnara with childhood friend Kartik Sarangmath and joined Y Combinator’s Summer 2025 batch. The company initially connected terminal-based coding agents to web and mobile interfaces, allowing developers to inspect activity, approve actions, and supervise sessions remotely. Its open-source platform subsequently expanded into managed agent infrastructure.

His approach to durable agents

• The log is the agent. Sehgal defines an agent by its append-only history of inputs, responses, tool calls, failures, and permissions. Like a saved video-game character, its identity survives changes in hardware and software when its history persists. Interfaces, model context, and audit trails become different views of that record.

• Serverless agent execution. Temporary workers reconstruct a session, advance it, record the result, and hand off to another worker. This architecture for serverless agents preserves pending approvals through crashes and treats machines as replaceable execution environments.

• Cross-model portability and log ownership. Sessions can branch across Claude, GPT, and Qwen without abandoning their history. Sehgal considers user-controlled logs essential because whoever controls that record controls an agent’s accumulated organizational knowledge, decisions, and permissions.

• Durability has boundaries. Compaction loses information, and replaying a session cannot reverse external actions such as sending email. His account of durable agent identity preserves the agent’s perspective without pretending to restore the entire world.