01 / POSITIONING

A RUNTIME.NOT ANOTHER LIBRARY.

Most agent frameworks are dependencies inside your application. IronCrew is a native runtime you deploy, operate, and audit — workflows in readable Lua, obligations enforced in Rust.

See how it works Talk to an engineer

02 / THE CATEGORY

MOST FRAMEWORKS ARE LIBRARIES. THIS IS A RUNTIME.

Both models are real engineering choices. They put the agent, its dependencies, and its controls in different places.

A library in your app

Agent logic iterates inside your service with your language's ecosystem and no extra deployable. It also inherits your application's dependencies, lifecycle, and blast radius.

A runtime you operate

Agent work runs on one inspectable surface with its own lifecycle. Journals, limits, human authority, provenance, and replica durability are runtime contracts, not application code.

03 / THE HONEST SPLIT

PICK BY SITUATION, NOT BY LOYALTY.

The split below is the honest one, not the flattering one.

A library wins when

You are prototyping at notebook speed and the loop is the product.

The workflow depends on your language's ecosystem — vector stores, loaders, application state.

One service should own the whole behavior, agents included.

You want the largest possible catalog of prebuilt integrations.

IronCrew wins when

A security review will ask for SBOMs, signatures, and source.

Production hosts must not carry a Python or Node runtime.

Human approval must fail closed instead of best effort.

Runs must survive restarts and coordinate across replicas on PostgreSQL.

Prompts must be able to stay on-prem with a local model.

IronCrew vs CrewAI IronCrew vs LangGraph

04 / WHAT YOU GIVE UP

THE COSTS, STATED UP FRONT.

Adopting a young runtime has costs. They are listed here instead of discovered later.

Smaller ecosystem

There is no integration marketplace. Reach comes from the HTTP tool, strict MCP servers, and custom Lua tools you write.

Lua, not Python

Flows are small and readable, but they are not written in your application's language. The dialect is small enough to read in one sitting.

Younger project

Fewer production miles than the incumbents. The release evidence and the published evaluation exist so you can judge the maturity yourself.

Judge it by the examples

05 / THE STACK

MOST STEPS DON'T NEED AN AGENT.

If a step is deterministic — integrations, extraction, transformation, schedules, retries — it belongs in IronFlow, IronCrew's sibling workflow engine. Reserve crews for the steps that need research, judgment, or human approval. An agent vendor telling you to use fewer agents is the honest version of this pitch.

See the escalation architecture Meet IronFlow

06 / EVIDENCE