Zero Ops - Agents Operate, Humans Govern

TL;DR

• The inversion: for decades humans have operated software systems — dashboards, metrics, alerts, automation, and lately Copilot — but the operator was always a human. Zero Ops makes the agent the operator and moves humans to governance.
• The definition: Zero Ops is not removing humans from operations; it's removing the operations from humans.
• The moat isn't the model: models are swappable and improve daily. What isn't swappable is your source code, production telemetry, institutional knowledge, and the skills and tools you hand the system.
• Maturity is crawl / walk / run / fly, and "fly" isn't merely execution — the agent acts, validates its own outcome, and hands the outcome to humans for review. That's outcome-based engineering: humans review outcomes, not every diff.
• Scale evidence: Microsoft has deployed 5,000+ agents internally on Azure SRE Agent, processing 1.5 million incidents, roughly 220,000 of them fully automated.

Key Points

A five-minute talk about compression (~00:41)

He opened by noting that "in agentic AI, compression is a thing" — he had five minutes to distill two years of building and running agents at Microsoft, and he intended to honor them. (He did.)

The status quo he wants to flip: for decades we've built dashboards, metrics, alerts, and automation, and more recently used Copilot to get work done — but the operator has always been a human.

Four conclusions on a single slide (~00:42)

1. What Zero Ops means: not removing humans from operations, but removing operations from humans — bumping humans up to higher-impact work like innovation and governance.
2. Models are swappable; context, governance, and tools are not. Models get better every day, so invest in the parts that don't turn over.
3. Models alone can't give you a useful agent. You need to govern it, control it, and feed it the right context. Enterprises love the intelligence they get — but if they can't control that intelligence, you won't earn their trust.
4. Trust requires verifiable metrics that let you validate what the agent did and what outcomes it produced.

Crawl, walk, run, fly — and what "fly" actually means (~00:43–00:44)

Zero Ops is a destination with a journey attached, and the last stage is the interesting one. "Fly" is not just the agent doing something — it's the agent doing something, validating the outcome itself, and providing that outcome for humans to validate.

His example: an API has a performance regression.

• Table stakes: the agent opens a PR. Opening a PR is easy.
• Fly: the agent opens the PR, deploys it to a test environment, validates that the regression is gone and performance is back to normal, and only then hands the result over.

That's outcome-based engineering — humans aren't reviewing every diff, they're reviewing the outcome.

Context is everything; governance is the entry ticket (~00:44–00:45)

• The model is not the moat. The moat is your source code, your production telemetry, your institutional knowledge, and the skills and tools you provide to the system. With those in place, agents can genuinely automate work end to end.
• To trust agents, an enterprise needs five things: to audit, control, evaluate, roll back with confidence, and prove with evidence that what the agent did was the right thing. Without these, enterprises simply won't go ahead.
• Permissions and metrics round it out: metrics validate the outcome, and without a permissions story enterprises won't deploy agents at scale.

Proof at scale: Azure SRE Agent (~00:45–00:46)

He stressed that none of this has to be built from the ground up — everything he described is baked into the product. Microsoft's internal numbers:

• 5,000+ agents deployed at enterprise level, used by some of the largest engineering teams at Microsoft.
• 1.5 million incidents processed.
• Roughly 220,000 incidents fully automated.

The closing formula: when context, governance, and metrics come together, that's when you achieve Zero Ops.

Quotes

"This is not about removing humans from operations. It's about removing the operations from humans." (~00:42)

The core definition of the whole five minutes.

"Enterprises love the intelligence they get out of it. But if they cannot control the intelligence, then you're not going to earn their trust." (~00:42)

The adoption bottleneck was never model capability — it's controllability.

"Humans are not reviewing every diff, they're reviewing the outcome." (~00:44)

Outcome-based engineering in one line.