Summary

Active Graph provides an event-sourced reactive graph runtime for building auditable, agentic systems. It offers a shared workspace paradigm where agent behaviors react to graph changes, ensuring every action is traceable, runs are resumable, and outcomes are forkable and diff-able, enhancing debuggability over traditional chat-based agents.

How It Works

The core is an event-sourced graph where mutations generate an append-only event log (the trace). Behaviors (function, class, or LLM-backed) subscribe to specific event types or graph patterns and propose state changes via patches. A novel "relation-behavior" primitive allows logic to be attached directly to typed edges, facilitating complex inter-object coordination. This event-driven architecture enables deterministic replay, forking, and structural diffing of system states.

Quick Start & Requirements

• Installation: pip install activegraph (core runtime + SQLite). Optional extras include [llm], [postgres], etc.
• Prerequisites: Python 3.11+.
• Quickstart: Run activegraph quickstart for fixture-based examples or activegraph quickstart --interactive for a 10-minute tutorial covering fork-and-diff.
• Documentation: Comprehensive docs available at docs.activegraph.ai.

Highlighted Details

• Event-Sourced Graph: Centralizes state and history, making all mutations auditable via an append-only event log.
• Fork-and-Diff: Branch any run at any event, reconfigure, and structurally diff results against the parent, leveraging cache replay for efficiency.
• Relation-Behaviors: Enables attaching logic directly to typed graph edges, a unique approach to agent coordination.
• Packs: Bundle domain-specific components (object types, behaviors, tools, prompts) for modularity.
• Per-Error Reference Pages: Each error links to detailed explanations, improving debugging.

Maintenance & Community

• Status: v1.0 (stable) released May 2026.
• Development: Described as in its "early public phase" with open issues and welcome documentation PRs. Code PRs are maintainer-only, requiring issue-first discussion. Roadmap items for v1.1 are tracked in CONTRACT.md.

Licensing & Compatibility

• License: Apache License 2.0.
• Compatibility: Permissive license suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The framework explicitly states it is not a chat, workflow, or traditional rules engine. While it supports optional Postgres or pluggable high-throughput backends, the default store is SQLite, positioning it away from production-grade graph databases. The system's integrity relies on well-defined behaviors; "bad behaviors produce bad graphs," though the runtime aims to make such issues inspectable. LLM tool use is noted as a v1.1 candidate feature.