The Log as the Primary Primitive
Most current agent architectures treat the model, runtime, or orchestration layer as the "agent." Ishaan Sehgal argues this is a mistake. Drawing an analogy to video games, he posits that the character is not the console or the game engine, but the save file. In AI, the agent is the session log—the append-only history of user inputs, model outputs, tool calls, and results.
By treating the log as the source of truth, the execution loop becomes disposable. A worker can pick up a log, advance the agent one step, and disappear, allowing any other worker to resume the session seamlessly. This shifts the agent from being a fragile, long-running process to a durable, stateful entity.
Emergent System Properties
When the log is promoted to a first-class citizen, several critical infrastructure properties emerge naturally rather than being bolted on:
- Reliability: Because the state is reconstructed from the log, if a process crashes (e.g., during a permission prompt), a new worker can resume exactly where the previous one left off.
- Scalability: Agents are no longer tied to a single machine or process. One worker can advance thousands of agents, eliminating the need for sticky sessions or complex state migration.
- Forkability & Multiplayer: Timelines can branch for parallel exploration (e.g., testing different models on the same history). Sharing an agent becomes a matter of granting access to the log, allowing for real-time observation and collaboration.
- Portability: Models and runtimes become interchangeable. Since the log is the identity, an agent can start on Claude, continue on GPT, and finish on Qwen without losing its history or context.
The Challenge of Log Ownership
Sehgal warns that the deepest form of vendor lock-in is not the model or the API, but log lock-in. As model providers (like Anthropic or Google) move toward "managed agents" that host the entire loop—memory, sandboxes, and compaction—they effectively own the agent. Because the log contains intimate company or personal data, ceding control of the log to a provider means ceding ownership of the agent itself.
Handling Compaction and External State
Two common objections to this architecture are addressed:
- Compaction: While logs grow indefinitely and context windows are finite, compaction is merely a projection of the log. It is a lossy, best-effort summary. By keeping the raw, durable log, you can always generate new projections, whereas discarding the raw log results in permanent data loss.
- External State: While agents may change external systems (e.g., sending an email or editing a file), the log is not intended to store the entire world. Like a game save file, its purpose is to store the agent's view of the world and the history of its actions, which is sufficient to maintain identity and continuity.