Skip to content

AgentStateGraph

AgentStateGraph is to agent state what Git was to source code — a content-addressed, branchable, blameable state primitive designed from the ground up for AI agents as the primary actor.

Every state change in AgentStateGraph is an atomic commit that carries the full provenance chain:

FieldQuestion
state_rootWhat changed?
intentWhy? (structured, queryable)
reasoningHow did the agent decide?
confidenceHow sure was it? (0.0–1.0)
agent_idWho did it?
authorityWho authorized it?
resolutionWhat was accomplished? Any deviations?

This is what git’s commit message tries to be — but structured, queryable, and enforced by the data model rather than by convention.

  • 73 MCP tools — any agent can connect immediately via stdio
  • O(1) branching and speculation — create, compare, and discard branches instantly
  • Schema-aware merge — CRDT-inspired conflict resolution
  • Epochs — sealable, tamper-evident audit bundles with Merkle root verification
  • Policy — authorization + cost-of-change gating with Cedar, Rego, and WASM evaluators
  • Taint & quarantine — mark sensitive paths, block writes, subscribe to changes
  • Namespaces — ref-layer isolation for multi-tenant deployments
  • Reminders — pull-based scheduling with priority, recurrence, and approval gating
  • Plans & tasks — shared task primitive with state machine, proofs, and agent assignment
  • 19 intent categories — structured vocabulary for agent reasoning
  • 4 storage backends — Memory, SQLite, Postgres (multi-tenant), IndexedDB (browser)
  • 6 language bindings — Rust, Python, TypeScript, Go, WASM, C FFI

Why it’s not git, a database, or an event store

Section titled “Why it’s not git, a database, or an event store”
ToolWhat it lacks
GitStructured intent, branching for non-text data, policy enforcement
Relational databaseBranching, blame, agent provenance
Event storeBranching, merge, speculation, policy gate
AgentStateGraphNothing — it’s purpose-built for this