Agents as Event Logs, Not Chat Threads

What it does

Active Graph is a Python runtime for long-running, stateful agents built around an event-sourced graph. Agents do not message each other directly; instead, reactive behaviors—functions, LLM calls, or logic attached to typed edges—fire in response to changes in a shared graph of objects and relations. Every mutation is an event appended to an immutable log, so the graph state is always a projection of its own history.

The interesting bit

The framework’s most unusual capability is fork-and-diff: branch any agent run at any event, reconfigure it, and structurally diff the result against the parent timeline without re-executing the shared prefix, because deterministic replay caches the common history. The README also treats failures as first-class behavior.failed events rather than exceptions, meaning crashes become audit trail entries instead of stack dumps.

Key highlights

• Behaviors subscribe to graph shapes using a Cypher subset plus event predicates, including “relation behaviors” that run on edges rather than nodes
• Strict replay mode re-fires every behavior and fails on divergence; permissive mode reconstructs state without re-execution
• Isolated EventSink workers stream live events without putting adapter I/O on the runtime hot path
• Cooperative bounded drains keep reads responsive during large derived-work drains via yielded quanta
• Core runtime has only two hard dependencies (click and pydantic); LLM providers and storage backends are opt-in extras

Verdict

Worth a look if you are building long-horizon agents where auditability and reproducibility matter. Skip it if you just need a lightweight LLM wrapper or a quick prompt chain.