yoheinakajima/activegraph

Active Graph is a Python framework for building long-running AI agent systems where every change is recorded and can be replayed later. Instead of having agents talk directly to each other, the framework sets up a shared workspace, called a graph, that holds objects and the typed connections between them. Pieces of code called behaviors watch the graph and react when something they care about changes. The README describes this as a group of agents sharing a workspace where everyone can see what was changed, by whom, and why. You install it with pip install activegraph and run activegraph quickstart to see a bundled example, called the Diligence pack, work against recorded fixtures. No API key is needed for that first run, and the output is meant to be the same every time. There is also an interactive version of the quickstart that scaffolds a behavior, runs it, and shows the fork-and-diff workflow. Optional installs add LLM providers (Anthropic, OpenAI), Postgres storage, or Prometheus metrics. Python 3.11 or newer is required. The core idea is event sourcing. Every mutation to the graph becomes an event in an append-only log, and the current state of the graph is treated as a projection of that log. Because the history is complete, a run can be resumed, forked at any point, or compared structurally against another run. The framework reuses cached LLM responses across the shared prefix of a fork, so branching does not trigger new model calls for work that was already done. Behaviors can be plain functions, classes, LLM-backed, or attached to typed edges between objects. They subscribe by event type, by a predicate, or by a subset of Cypher patterns for matching shapes in the graph. Failures of a behavior are themselves logged as events rather than raised as exceptions, so the audit trail stays complete.