Running IT for 200+ locations with supervised AI agents.
For two years I was the IT function for a five-brand, 200-plus-location franchise group — a lean team and a private-equity budget, against an estate most companies staff with a department. The way that math works is supervised AI agents. Here is how I built it, and what I would tell anyone trying the same.
The constraint shaped the method
Five brands, 200-plus locations, five Microsoft tenants, and a budget that did not include hiring a platform team. I could not out-staff the problem, so I built capacity instead: production systems built with supervised AI agents. The agent I used is Code Puppy, the open-source coding agent adopted at Walmart, where it won the company's President's Innovation Award. I did not build the agent — I built with it, under tight supervision.
Supervision is the product, not the model
The interesting part of agent work is not the model — it is the system around it. I ran a planner agent that orchestrated specialists: a crawler, a programmer, a code-reviewer, a security-auditor. Every destructive action waited behind a human-in-the-loop gate, and every release had to clear an evaluation harness — a judge — before it shipped. Control Tower, the governance plane across the five tenants, ships behind 7,386 automated tests and a 48/48 judge score. The judge caught regressions the test suite missed and quietly became the real release gate. If you take one thing from this: stand up the eval harness first.
Identity is the filter, not an afterthought
Knowledge Fabric answers support questions drawn from more than 100,000 help-desk tickets, plus point-of-sale configs and SharePoint. The hard part was never search — it was that answers have to respect brand and role boundaries. Staff in one brand should never surface another brand's playbook. So it is grounded on Entra ID — 1,523 verified identities across a 15-slot attribute schema — and identity is the filter on every retrieval, not a check bolted on later. Phase one closed at 52/52 QA checks.
You cannot govern what you cannot see
Estate Trace started from a question no one could answer: what do we run, and what does it cost — across 41 repos, 11 subscriptions, roughly 196 Azure resources, and reseller-billed licensing. It pulls all of that, read-only, into a DuckDB warehouse and one canonical workbook, and 12 of 12 verification checks have to pass before that workbook is published as source of truth. It quantified the IT run-rate and a roughly 40% optimization path with 25 prioritized actions. Once people trusted the numbers, every cost conversation started there.
Zero stored secrets
None of this stores a credential. Automation authenticates to Azure with OIDC workload identity federation across all five tenants — no client secrets in repos, no API keys in CI, no rotation nightmares. For a multi-tenant estate, removing an entire class of secret-handling risk was worth more than the hours it took to wire.
What I would tell you
Supervise, do not trust — the gates are the point. Ground on identity from day one; retrofitting role boundaries later is painful. And measure before you optimize. I left the group before the framework was rolled out portfolio-wide, which is exactly why I am open-sourcing the governance layer as TenantFleet — the patterns worked, and other franchise operators, MSPs, and PE-backed teams should not have to rebuild them from scratch.
The through-line is the same as every role I have had: understand what people actually need, then remove the technological obstacles that make their day harder. Agents just let a small team do that at a scale that used to require a big one.
Building something that needs this kind of system? Let's talk — or see the case studies.