The Log is the Agent reframes agent runtime around event sourcing

Most agent frameworks start with the language-model loop, then add tools, rules, memory, and logging. Yohei Nakajima’s paper The Log is the Agent proposes the reverse. In ActiveGraph, an append-only event log is the source of truth. The working graph is a deterministic projection of that log, and behaviors react to graph changes by emitting new events.

The payoff is operational rather than cosmetic. If the log is complete, a run can be replayed deterministically. A session can be forked at a specific event without re-running the shared prefix. Artifacts can be traced from a high-level goal down to the model call or routine that produced them. That makes the design closer to event sourcing and durable workflow systems than to conventional chat memory.

The Hacker News thread quickly moved into implementation comparisons. Several commenters noted that their own harnesses store event graphs or session logs in similar ways. Others pushed on the familiar pattern of agent developers rediscovering the same primitives independently. That reaction is useful context: the paper is not interesting because “logging” is new, but because it treats the log as the runtime substrate rather than a debugging accessory.

The paper is careful about its claims around self-improving agents, describing suitability rather than demonstrating a solved system. Still, deterministic replay, fork-and-diff, and lineage tracing are concrete capabilities that current agent tooling needs. When an agent run fails, teams want to reproduce it, branch from the failure point, and understand which call produced which artifact. An event-sourced graph makes those questions first-class.