Modernization track · Advanced · self-paced ~14h with labs (2 days as a cohort) · prerequisite: CS520 or CS530, and an MCP-capable client (Claude Code, Kiro, Copilot agent mode or any MCP client)
Wire a coding agent to the SteelFrame MCP and run the understand → convert → run both → compare loop with the agent doing the typing and the oracle doing the judging: what assistants do well (explain, inventory, scaffold, tests) and badly (judge equivalence); the four tool families taught as methods, not products; MCP tools, scopes, workspaces and the audit trail; the governed loop in the shop’s chosen style; the five failure sites as drills — packed rounding, EBCDIC mismatch, REDEFINES overlay, control-flow fall-through, a hallucinated business rule; and the evidence trail an auditor accepts. 24 short videos (about 3h) on CardDemo via MCP.
Self-paced available Cohort
Open for enrollment now. The lesson videos for this course are in production — release order across the track: CS510 → CS520 → CS530 → CS521 → CS535 → CS500 → CS550 → CS525 → CS540 → CS545 — and every session below is published as its lessons and lab now and marked Coming soon until its videos are released.
Plans: Free individual at no cost, or Paid individual at $349 per course, one-time — code review, extra exercises, grading and priority support.
Every session ends in a lab on your own environment. Labs are machine-checked against expected outputs — you and your L&D team always know exactly where you stand.
Coming soon The videos for this session are being produced; its lessons and lab are published here now.
Four videos. Explain — good at structure, silent about the rule that matters, checked against CBTRN02C in BROWSE; inventory — fast and only as good as the tool it calls, inventory_scan beside a confident guess; tests — scaffolding is cheap and the expected value must come from the oracle, a ZWSUNIT deck failing on an invented count; it cannot judge equivalence — a public AI-generated Java CardDemo verified against the real COTRN02 ladder.
LAB Run the course verdict table on the public AI-generated build for five scenarios and report every difference against the live CardDemo run.
Coming soon The videos for this session are being produced; its lessons and lab are published here now.
Four videos, each walked on SteelFrame artefacts with your own client. Understand → refactor-into-services → transform → validate with generated tests (the WCA4Z/Bob shape) on COACTVWC; assess → extract rules → requirements → traceable code (the AWS reimagine shape) on COTRN02C’s running ladder; assessment → generated code → dual run (the Google shape) as two SteelFrame runs beside one Java run; open agent pipelines turning INTCALC into a Spring Batch job — with the oracle gate between every stage and what each leaves for the human.
LAB Run the open pipeline on TRANREPT with the oracle gate inserted; submit the gate’s verdict and the three-way dual-run table.
Coming soon The videos for this session are being produced; its lessons and lab are published here now.
Four videos. The tool surface and the six-call loop — create_workspace, upload_source, compile_program, run_program, decode_records, export_state — run against the live engine; scoped zws_ tokens, what the agent cannot do and the audit log that proves it; connecting a client and the first whoami; the governed-scope gate.
LAB Gate: your agent runs the full loop on a program in your workspace under a scoped token, attempts one out-of-scope write and is refused; transcript and audit lines as evidence; workspace destroyed.
Coming soon The videos for this session are being produced; its lessons and lab are published here now.
Four videos. Golden first, agent converts in the shop’s style, suite judges, agent fixes only what the suite names, human reviews declarations; prompt and tool-description hygiene — the sloppy prompt that tries to "fix" the golden and is refused; recording evidence — transcript, call log, capture ids, suite reports, sign-off; the conversion gate.
LAB Gate: CBTRN03C converted by your agent in your chosen style, suite green against your TRANREPT golden, evidence pack complete.
Coming soon The videos for this session are being produced; its lessons and lab are published here now.
Four videos. Packed rounding and EBCDIC mismatch shown in the bytes on SteelFrame before the catch; REDEFINES overlay and control-flow fall-through shown in EDIT before the wrong counts; the hallucinated business rule — a minimum amount the COBOL never had — against the real CT02 screen; the hallucination gate.
LAB Gate: a provided agent transcript and build with one hallucinated behaviour — found with the suite against your goldens, the rule and the transcript line where it was introduced named.
Coming soon The videos for this session are being produced; its lessons and lab are published here now.
Four videos. The brief and the golden captured on SteelFrame; the governed loop live — red, red, green; the evidence reviewed as the human who signs, capture ids and audit lines verified; the certificate.
LAB Capstone: an unseen CardDemo program converted in your chosen style by your agent loop, suite green against your own golden, the evidence pack self-reviewed with the checklist.
CobolStack Certificate — AI-Assisted Modernization (CS550) is issued from the lab harness’s own record of your runs — not from attendance. To earn it you must clear every gate below:
The certificate attests that you can put an AI agent to work on a real mainframe estate under governance — scoped, audited, oracle-judged — and produce the evidence that separates assisted modernization from guesswork.
Your seat is a private, full z/OS-compatible environment exposed to your AI agent through the SteelFrame MCP (scoped tokens, workspaces, the full tool surface) beside the AWS CardDemo estate; your own MCP client and Java toolchain in the browser — nothing to install, snapshot and reset per exercise. By the end of the course the lab harness will have verified: three lab gates — a governed token scope demonstrated with the refused out-of-scope write in the audit log, an agent-converted CardDemo program proven equivalent by the suite against your own golden, and a seeded hallucinated behaviour caught by the suite and traced to its transcript line — and the capstone: an unseen CardDemo program converted in the chosen style by your agent loop, oracle-judged, with the transcript and capture ids as evidence.