01 / THE STACK
IronFlow runs the process. IronCrew handles the judgment. If you already know what should happen, automate it deterministically; when the system must determine what should happen, hand that step to a crew — and only that step.
02 / THE RULE
If you already know what should happen, use IronFlow. If the system needs to determine what should happen, use IronCrew.
Integrations, extraction, transformation, schedules, retries, routing, and persistence. A deterministic DAG engine with 102 built-in nodes, in the same shape as IronCrew: one Rust binary, flows in Lua.
Meet IronFlowSpecialist agents research, challenge, and decide. Humans answer questions and approve consequences. The crew returns a structured decision — it does not wander off and act.
See how it works03 / ESCALATION
A crew buys measured quality at a measured cost — +13.3 points grounded correctness at 3.2x the tokens in the published evaluation. Spend that where ambiguity is expensive, not on every case.
IronFlow
Webhooks, schedules, files, and API calls enter one deterministic pipeline.
IronFlow
Extraction, validation, transformation, and routing run without any model involved.
IronFlow llm node
Plain classification and extraction use a single model call inside the flow.
IronCrew
High-value, uncertain cases go to a crew: researcher, analyst, critic — and a structured decision comes back.
IronCrew HITL
When the decision is consequential, an approval gate holds fail-closed until a person signs off.
IronFlow
The decision returns to the deterministic pipeline: update systems, notify, persist, audit.
04 / ROUTING
IronFlow has its own llm node — simple semantic steps do not need a crew. The overlap is deliberate and stated, not hidden.
| Requirement | Runs on |
|---|---|
| Call an API, move data, run SQL | IronFlow |
| Extract a PDF, Word, or Excel file | IronFlow |
| Chunk, embed, and store documents | IronFlow |
| Schedule, retry, branch, notify | IronFlow |
| Simple classification or extraction | IronFlow llm node |
| Multi-perspective analysis under ambiguity | IronCrew |
| Research, challenge, and synthesize | IronCrew |
| Ask a human for missing context | IronCrew |
| Approve an AI-proposed action | IronCrew HITL |
| Execute the approved business action | IronFlow |
05 / PATTERNS
Concrete shapes, not industries. Each one keeps the deterministic majority cheap and reserves the crew for judgment.
Intake → deterministic checks → uncertain cases to a crew → confident results proceed, sensitive ones wait for a human.
Fraud alerts, claims, compliance exceptions, support escalations.
Extract and normalize with IronFlow → analyst, specialist, and reviewer agents read → structured findings → systems updated.
Contracts, procurement, regulatory submissions, due diligence.
Alert arrives → logs and metrics collected deterministically → a crew forms a root-cause hypothesis → remediation waits for approval.
Platform operations, security events, quality incidents.
Scheduled collection and deduplication → researcher, skeptic, and writer agents → an intelligence brief lands where the team reads.
Market and competitor monitoring, regulatory watch, technology scouting.
06 / GOVERNANCE
Agents propose structured intent. Schema validation, fail-closed approval gates, and human sign-off sit between proposal and execution. The deterministic engine performs the action.
INTEGRATION TODAY
IRONCREW REST + SSE / IRONFLOW WEBHOOKS + HTTP — PLAIN HTTP ON BOTH SIDESRead the HITL contractBUILD THE STACK