Where Flow proved the pattern
Mainframe engineers work across ISPF, terminals, spool viewers, and job schedulers. Flow's first use case collapsed that into one canvas with local AI at every interpretation point.
Submit-and-interpret, with and without Flow
Without
- Open ISPF, navigate to the JCL.
- Submit the JCL manually.
- Check the job queue periodically.
- Open the spool viewer.
- Copy results to email for QA.
- If failed, re-edit or ask a senior dev.
With Flow
- Drag nodes onto the canvas (or generate from a prompt).
- AI Syntax Validator checks the JCL locally, with zero network.
- The engine submits the job with an authenticated profile.
- Download Spool retrieves output. The Spool Interpreter annotates it in plain language.
- Failure recovery suggests a fix and can re-run on approval.
- An email notification with the AI-generated summary goes to QA.
All AI inference happens on the engineer's machine. No data leaves the laptop.
Five steps, two of them thinking
The flow submits the job through the authenticated CLI, validates the syntax with a local model, downloads the spool, interprets it with a second local model, and sends the summary. When a step fails, the engine passes the failure context back to the models for diagnosis. The retry state lives in the orchestrator, never in the model.
Spool from a failed job can also seed flow generation. You feed the spool to the generator and get a proposed recovery flow to review.
flow "JCL Pipeline"
submit[action: "Submit JCL"] {
command: "zowe jobs submit local-file"
retry: 2
timeout: 30s
}
validate[ai: "jcl-validator:v1"] {
model: "jcl-validator:v1"
}
interpret[ai: "spool-interpreter:v1"] {
model: "spool-interpreter:v1"
}
email[utility: "Send Email"] {
template: "job-failure-summary"
}
submit --> validate
validate.success --> interpret
interpret --> email when status == "failed"Mainframe flows in production shape
ACF2 security scan, without the mainframe tax
Michal, systems programmer at a Tier-1 global bank
Turn an hour-long, four-system relay into one click - and make 'your toolkit was out of date' impossible to hit by accident.
Four tools and ~12 manual steps become one Play button
The code-validate-run loop
A developer who just edited a JCL job
Stop typing four CLI commands and waiting between each - wire the loop once, run it on every edit.
Four commands and three waits become one click per edit
Scheduled, fleet-wide scans
A batch operator responsible for many LPARs
Run the same vetted scan across the whole estate on a schedule - and never let one back-level LPAR rot silently.
Manual relay times N LPARs becomes a schedule that runs itself
The platform is not mainframe-specific
The same architecture extends to CI/CD and QA. A new domain means new models, not a new platform.