Introduction

AgentStateGraph is a content-addressed, versioned, branchable structured state store designed as an infrastructure primitive for intent-based systems.

The Problem

AI agents don’t execute linear scripts — they explore state spaces. An agent asked to “set up a cluster for ML training” tries different approaches, compares outcomes, and picks winners. It needs to:

• Branch to try approaches without risk
• Compare outcomes side-by-side
• Merge the winner back
• Report what was done, what deviated, and why
• Record the full reasoning chain for audit

No existing tool supports this natively. Git is text-oriented. Databases lack branching. Event sourcing is append-only.

What AgentStateGraph Provides

Every state change in AgentStateGraph captures the full provenance chain:

Field	Question
state_root	What changed?
intent	Why?
reasoning	How did the agent decide?
confidence	How sure was it?
agent_id	Who did it?
authority	Who authorized it?
resolution	What was accomplished? Deviations?

Key Features

• Content-addressed Merkle DAG — immutable, deduplicated history
• Schema-aware merge — CRDT-inspired conflict resolution
• Speculative execution — O(1) branching, instant discard
• Multi-agent orchestration — scoped sessions, delegation, intent trees
• Epochs — sealable, tamper-evident audit bundles
• 73 MCP tools — any agent can connect immediately
• HTTP REST API — 22 endpoints with CORS, run with --http
• Browser explorer — interactive data viewer at agentstategraph.dev/explorer/
• 6 language bindings — Rust, Python, TypeScript, Go, WASM, C FFI
• 4 storage backends — Memory, SQLite, Postgres (multi-tenant), IndexedDB (browser)
• 14 crates — modular core, storage, MCP, policy, taint, tasks, reminders, and bindings
• Plans & Tasks — shared agentstategraph-tasks primitive with state machine, proofs, blockers, agent assignment
• Policy — authorization + cost-of-change gating with Cedar, Rego, and WASM evaluators; Ed25519 signing
• Taint & quarantine — agentstategraph-taint mark-and-sweep enforced at commit time
• Namespaces — ref-layer isolation for multi-project / multi-tenant deployments
• Reminders — pull-based scheduling with priority, recurrence, and approval gating
• 19 intent categories — Explore, Refine, Fix, Rollback, Checkpoint, Merge, Migrate, Plan, Taint, Untaint, Quarantine, Unquarantine, Watch, Unwatch, PolicyPropose, PolicyRatify, PolicySupersede, PolicySign, Custom
• Schema migrations — /_meta/schema_version guard + agentstategraph-migrate registry + agentstategraph-mcp migrate CLI