Compare · frameworks
Not a competitor. A different layer.
LangGraph, CrewAI, the Claude Agent SDK, the OpenAI Agents SDK, Google ADK, Mastra, Agno — these decide how an agent thinks. Velone decides what happens when it acts. You keep the one you have.
01 / credit where it is due
Replacing your framework is a migration nobody asked for.
LangGraph's durable graph execution, checkpointing, and time-travel are excellent, and LangSmith is a genuinely useful tracing product. CrewAI's role-based model maps cleanly onto how people describe work. The Claude and OpenAI SDKs are the fastest path to a working loop. None of that is something we want to rebuild.
So Velone ships adapters instead of an opinion. Your graph, your crew, your prompt engineering, your evals — unchanged. Only the execution path moves.
02 / where it stops
The gap we exist to fill.
- Isolation
- A framework runs tools in its own process. Untrusted code executes with the privileges of whatever launched it, which in practice is a developer laptop or a CI runner holding real credentials.
- Identity
- Frameworks model an agent as an object in memory, not a principal your identity provider knows about. There is no durable name to attach an authorisation to.
- Enforcement
- Guardrails and callbacks run inside the same process the model influences. They are useful for quality and hopeless as a security boundary, because anything in-process can be routed around.
- Audit
- Traces are built for debugging: sampled, mutable, retained for weeks. An auditor needs complete, immutable, and attributable.
- Secrets
- The standard pattern is an environment variable in the process running the loop. Every tool and dependency in that process can read it.
- Human approval
- Several frameworks have an interrupt primitive, which is the right shape. What is missing is a signed authorisation bound to a named person and carried in a record.
03 / side by side
| Velone | Agent framework | |
|---|---|---|
| Decides how the agent reasons | No | Yes |
| Graph, crew, or loop abstractions | No | Yes |
| Prompt and eval tooling | No | Yes, often |
| Hardware-isolated execution | Yes | No |
| Durable principal identity | Yes | No |
| Out-of-process policy gate | Yes | No |
| Zero standing credentials | Yes | No |
| Tamper-evident evidence | Yes | No |
| Default-deny egress | Yes | No |
| Works with the other layer | Yes, by design | Yes, via adapters |
04 / choosing
Pick the right one.
choose them when
- You are deciding how to structure an agent's reasoning — that is their job, not ours
- You need checkpointing and time-travel over a graph
- You want prompt management, datasets, and evals
- The agent is read-only and touches nothing that matters
choose velone when
- You have a working agent and the blocker is security, not capability
- The agent executes code, moves money, or writes to customer records
- You need the same boundary across two frameworks and three models
- Someone has asked how you would prove what the agent did
05 / questions
Do we have to drop LangGraph?
No. The adapter is a checkpointer plus a tool-node wrapper, and Velone sessions map one-to-one onto LangGraph threads. Your graph definition does not change.
Is this just a wrapper around tool calls?
At the integration point, deliberately yes — that is what makes adoption cheap. What is behind the wrapper is a microVM fleet, a policy engine, a credential broker, and an audit chain. The thin seam is the feature.
What about LangSmith or Braintrust?
Keep them. They answer 'is the agent any good', which is a different question from 'can it do this and can you prove it did'. The two records serve different readers and do not conflict.
Will frameworks add this themselves?
They will add more guardrails, and those will keep running in-process where the model can influence them. A credible boundary has to live outside the process that the model shapes, which is an infrastructure product rather than a library feature.
keep reading
Ring 0
Keep the framework. Add the kernel.
Tell us which harness you run. The integration is usually an afternoon, and the policy conversation is the part worth your time.