The Veridex Agent Fabric
The only TypeScript agent runtime that gives you safe, stateful, long-context autonomy without the framework magic.
Veridex is a TypeScript-first agent runtime designed around a single thesis: every consequential thing an agent does must be typed, gated, observable, and resumable. We didn't add safety as a feature; we built the loop around it.
Why another agent framework?
Every team building agents in production hits the same wall:
| Pain | Where it bites | What competitors do | What Veridex does |
|---|---|---|---|
| The framework rewrites prompts behind your back | Debugging is impossible | LangChain/LangGraph mutate state in reducers | Event-sourced loop — every transition is a typed event you can read (ADR-0049) |
| Long conversations degrade | Quality cliff after ~30k tokens | Buffer + summarise on overflow | Strict effective window with per-section budgets and homeostatic compression (ADR-0050) |
| "Memory" is just transcript replay | Cost bloat, staleness, drift | Vector store as a tool | Four tiers + reconciler — promotion, decay, conflict resolution (ADR-0051) |
| Tools are functions with API keys | TPA, prompt injection, traversal | "Trust the developer" | Typed contracts + sandbox + safety classes (ADR-0052) |
| No human-in-the-loop without rewriting | Treasury, healthcare, legal | External workflow engine | Approvals + checkpoint/resume as a first-class loop primitive (ADR-0054) |
| "Ship on vibes" | Production regressions | Manual prompt eval | Stateful eval harness with golden traces and red-team corpus (ADR-0062) |
What you get
@veridex/agents— the core runtime: tools, policy, memory, checkpoints, transports, evals. Zero magic; one well-documented loop.@veridex/agents-react— focused hooks:useRun,useApprovals,useTrace,useContextHealth,useMemory,useSkills,useBudget.@veridex/agents-adapters— incremental migration from LangChain, LangGraph, OpenAI Agents SDK, OpenAPI, PydanticAI.@veridex/agents-openclaw— OpenClaw / Pi-style context files and skills.@veridex/agents-control-plane— multi-tenancy, policy packs, approval workflows, durable trace storage.@veridex/agent-security— framework-agnostic security gateway, packs (injection, tool poisoning, secrets, endpoint allowlist, budget, shell safety, handoff), and adapters for LangChain and CrewAI.@veridex/agents-treasury— money-movement capability layer: idempotency, ceilings, time-lock, sanctions, signed evidence bundles, red-team eval suite.
Quick comparison
| Dimension | Veridex | LangChain / LangGraph | OpenAI Agents SDK | Mastra |
|---|---|---|---|---|
| Event-sourced loop | ✅ first-class | ❌ opaque reducers | ❌ black box | partial |
| Strict context budget | ✅ enforced per-turn | ❌ overflow → summarise | ❌ raw window | ❌ |
| Tiered memory + reconciler | ✅ 4 tiers, lifecycle | partial (vector add-on) | ❌ minimal | partial |
| Tool sandbox + safety classes | ✅ default | ❌ | ❌ | partial |
| Policy engine | ✅ priority-ordered rules | ❌ ad-hoc | ❌ | partial |
| Human approvals + checkpoint/resume | ✅ first-class | bolt-on | ❌ | partial |
| Signed audit / evidence bundles | ✅ HMAC + Ed25519 | ❌ | ❌ | ❌ |
| Stateful eval + red-team corpus | ✅ built-in | external | external | partial |
| Multi-tenant control plane | ✅ opt-in package | external | ❌ | ❌ |
Start here
- Quick Start — five minutes to a production-shaped agent.
- Context Compilation — the engine that keeps long-running agents coherent.
- Tools — typed contracts and safety classes.
- Policy — priority-ordered, replayable rules.
- Approvals — humans in the loop, durably.
First principles, not framework magic
Every choice in this docs site is derived from a concrete failure mode of existing frameworks and a concrete piece of research. We cite both. If you ever wonder "why is Veridex doing it this way?", click through to the relevant ADR.