Two ways: long jobs, overnight reviews, migrations, research, that run for hours with receipts, and company context for your coding agents over MCP, which ends the token-burning rediscovery of your own decisions.
'Review all 45 backend routers overnight' is a real Heidi job: hours of work, live progress, the plan in your inbox by morning, every step logged.
And the 'why' questions, why is the service split, why is the cap 3, get answered from memory with receipts instead of from a repo crawl. Engineers keep the judgment; Heidi carries the archaeology.