← back
The Log Is The Agent - Ishaan Sehgal, Omnara
Takeaway
Make the event log primary so agent identity survives worker failures and runtime changes.
Summary
- An agent’s durable identity is its append-only log of inputs, outputs, tool results, permissions, failures, and state transitions.
- Workers can reconstruct state, advance a session, append results, and disappear, enabling recovery and flexible scheduling.
- Context, UI, auditing, and compaction are projections of the log; retaining raw history preserves information lost by summaries.
- A log records the agent’s view of external actions but cannot undo emails or make the outside world deterministic.
- This design supports session forking, sharing, and migration across models or runtimes.
event-sourcingagent-statedurability
Original description
Think about a character you've spent 100 hours playing in a video game like Skyrim or Elden Ring. What exactly is your character? Is it the game engine (the loop)? No. Is it the PlayStation console (the compute)? No. Is it the controller (the tools)? No. Your character is the save file (the data). If your PlayStation bursts into flames, your character isn't dead. You just buy a new PlayStation, download your save file from the cloud, and your character is exactly where they were, mid-swing. The identity, history, and current state of your character live entirely in the data. Today, most attention goes toward frameworks, orchestration layers, context engineering, specs, and tools. But as models become more capable and generally intelligent, the differentiator shifts away from these abstractions and toward the underlying infrastructure. The Engine becomes interchangeable. The Brain (LLM) commoditizes. The Hands (tools) are just APIs. What actually persists across models, runtimes, and machines is the session log. That’s where continuity, identity, and state actually live. A decade ago, Martin Kleppmann argued that databases should be understood as projections over an append-only log. I think the same thing is now happening with agents. Today, agents are treated like complicated black boxes made of loops, models, and tool calls. But at its core, the agent is the session log (the save file). Everything else is swappable. Once the log becomes the primitive, entirely new system properties emerge: durability — agents survive crashes, disconnects, and machine failure continuity — sessions can be resumed from anywhere, on any device forkability — timelines can branch for parallel execution and exploration addressability — agents become durable entities that can be referenced and revisited observability — execution can be monitored and steered in real time portability — models, runtimes, and machines become interchangeable But this creates a new infrastructure question: If the model provider owns the log, does the model provider own the agent? This talk explores why the future of agent infrastructure isn't what we're focused on today, but rather durable, portable logs that make agents persistent across models, runtimes, and machines. Speakers: Ishaan Sehgal (Omnara): Ishaan Sehgal is the CEO and cofounder of Omnara (YC S25), where he builds managed infrastructure for AI agents. X/Twitter: https://x.com/ishaansehgal LinkedIn: / ishaan-sehgal GitHub: https://github.com/ishaansehgal99