Veridex vs LangGraph

LangGraph is LangChain's stateful graph runner. It's a real improvement over AgentExecutor for explicit control flow. We still find it short for production agents:

Dimension	LangGraph	Veridex
State model	User-defined reducer over a `StateGraph`	Event-sourced; typed `TraceEvent` stream; deterministic replay
Checkpoints	`checkpoint_id` per state snapshot	Differential checkpoints + content-hashed event log + signed bundles
Human-in-the-loop	`interrupt_before/after` + manual resume	First-class `escalate` verdict; `ApprovalManager`; workflow tiers
Tools	LangChain `Tool`	Typed contract + safety class + sandbox
Context budget	Manual	Strict $V_e$ , per-section budgets, compression-as-steady-state
Policy	If/branch in graph	Priority-ordered rules; pack composition; staged rollouts
Adversarial proofs	External	Built-in red-team suite, runs in CI
Multi-tenancy	Out of scope	Control plane (opt-in)

They compose

A LangGraph graph can call a Veridex agent for a high-stakes subgraph; the rest of the graph keeps running.

import { veridexNode } from '@veridex/agents-adapters/langgraph';
 
const graph = new StateGraph(MyState)
  .addNode('research', researchNode)              // stays in LangGraph
  .addNode('treasury', veridexNode({              // governed by Veridex
    agent: treasuryAgent,
    mapState: s => s.userQuery,
    mergeResult: (s, r) => ({ ...s, paid: r.output }),
  }))
  .addEdge('research', 'treasury');

The treasury node runs under Veridex's policy + approvals + evidence bundling. The rest of the graph keeps its LangGraph semantics.

When LangGraph is enough

The graph state is small and well-typed.
HITL is rare and short.
Audit is informal.
Single-tenant deployment.

When Veridex is the right tool

Treasury, healthcare, compliance.
Hours-to-days workflows with multi-tier approvals.
Adversarial threat models.
Multi-tenant fleets with per-tenant governance.

vs LangChain vs LlamaIndex