Ishaan Sehgal proposes treating an agent’s append-only event history as its durable identity, rather than tying the session to a particular model or worker. He describes rebuilding state from eventsEvent sourcing records application state changes as an ordered sequence of events, allowing state to be reconstructed by applying those events., running a step and saving the result so another worker can continue.
He describes compaction as a lossy projection of the full history and distinguishes the agent’s recorded view from external state. Rewinding a log, for example, cannot undo an email that was already sent.
Sehgal argues that log-centered design can support recovery, forking, collaboration and migration between runtimes. He also discusses ownership of sensitive session historyData privacy governs how personal, confidential, or sensitive information is collected, used, shared, retained, and protected in AI systems. and presents Omnara’s planned managed-agent architecture; these are design claims rather than guarantees.
Watch on YouTube
