The situation
Mainframe modernisation planning is usually slow because the work is specialist, sequential, and expensive to get wrong. The organisation expected a year of analysis before it would have a plan it was prepared to stand behind. The constraint was not willingness. It was the time required to read the estate, sequence the work, and get humans to approve the result.
Agent roles and orchestration
The system did not ask one general assistant to “plan a modernisation”. Specialist agents worked a pipeline:
- A planner framed scope, dependencies, and the order of investigation.
- Specialist assistants analysed domains that a general model treats poorly when left alone.
- An architect role turned findings into an options set rather than a single unchallenged design.
- Documentation agents produced artefacts humans could review, rather than a chat transcript.
Orchestration meant shared context and explicit handoffs. One agent’s output was another agent’s input, not a pile of disconnected prompts.
Validation, governance, and limits
Autonomy stopped at the gate. Humans validated analysis, approved sequencing, and remained accountable for the plan. Agents did not release a modernisation programme. They compressed the work required for a human to make a defensible decision.
Organisational implication
This is the AI-SDLC point. The gain was not “AI writes COBOL faster”. The gain was a governed, agent-initiated planning workflow with human approval, measured as planning-time improvement. That is the same operating idea the Copilot programme needed at organisational scale: assistants become a system only when roles, gates, and measurement exist.
Public facts are limited to what can be supported without client identification. No additional metrics are claimed here. Validation is available under NDA after fit confirmation.