• Towards AGI
  • Posts
  • Microsoft Has 50 Million AI Agents On Its Books. Can You Count Yours?

Microsoft Has 50 Million AI Agents On Its Books. Can You Count Yours?

here's what ignite means for your agent roadmap

Today, we're diving into:

  • Gen AI: Microsoft says it already has more than 50 million agents in Agent 365, and it wants to be the control plane for all of them

  • Hot Tea: A new US rule makes data centers anywhere in the world need a license if their AI chips might serve restricted countries

  • OpenAI: OpenAI is letting four outside safety groups inside to test whether its models could quietly work against it

Microsoft Has 50 Million AI Agents On Its Books. Can You Count Yours?

Microsoft used its Ignite keynote in San Francisco to make one big claim: it wants to run the operating system for AI at work. In the edited keynote published on November 17, Satya Nadella said Microsoft already has "over 50 million agents in Agent 365."

Agent 365 is the control plane. Microsoft describes it as extending your existing governance, identity, security and management systems to the agents running in your enterprise. Nadella called the new Agent 365 Business Analytics "your ERP system for AI usage."

Fifty Million Agents. How Many Of Yours Have An Owner?

Find out whether your data is ready for agents that never clock off.

An executive overseeing thousands of AI agents from one control console

Microsoft's new Autopilot is built to keep work moving around the clock, and people want the same thing outside the office: an agent that does the legwork for them. That is what OpenMap does, and it is now live: from shopping to planning a group trip, ask in your own words and your agent does the legwork. Try it free at openmap.ai.

OpenMap is live. Shopping, group trips and more: ask in your own words and your agent does the legwork. Try openmap.ai.

What Microsoft Actually Shipped

Three announcements matter most if you run AI at enterprise scale.

  • Agent 365: one place to govern, secure and track the agents across your company, now with business analytics and cost management.

  • Microsoft IQ: what Constellation Research calls "a new enterprise context management system," pulling together Work IQ, Fabric IQ, Foundry IQ and Web IQ with 350 third-party connectors.

  • Autopilot: an autonomous agent that Microsoft says "keeps moving your work forward 24×7."

Read together, the message is clear. Agents are no longer something you pilot. They are something you operate, around the clock, at volume.

Even Microsoft Admits The Blind Spot

"Today, we can see almost everything models touch. But we cannot see what happens inside the model." Satya Nadella, CEO, Microsoft

Nadella went further, calling these "nested black boxes" simply unacceptable. When the company selling the control plane says visibility is the open problem, you should take it seriously.

This Is Not An Agent Problem. It Is A Context Problem.

The wrong reaction is to count your agents, buy a dashboard and call it governance. A dashboard can tell you an agent ran. It cannot tell you whether the data it read was right.

Notice what Microsoft paired with the control plane: a context system. That is not an accident. An agent working 24×7 makes thousands of small decisions from whatever definitions, documents and records it can reach.

An AI agent drawing on enterprise data silos joined into one context layer

The real bottleneck is not how many agents you can launch. It is whether every one of them reads from context your business has agreed is true, with an owner you can call when it is not.

The better alternative is to build that context layer once and let every agent, from any vendor, draw on it. This fundamentally changes the question from "which agent platform wins?" to "whose definitions do my agents trust?"

What You Should Do This Quarter

  • Inventory every agent. Include the ones teams built in Copilot Studio, Foundry or elsewhere without telling IT.

  • Map what each agent reads. Write down every source, its owner and when it was last checked.

  • Price agents like headcount. If Microsoft calls Agent 365 analytics an ERP for AI usage, your finance team should see those numbers too.

  • Stay vendor neutral on context. Your definitions should outlive any one agent platform.

Agents you can trust start with trusted enterprise context. DataManagement.AI helps organizations create that foundation by unifying metadata, lineage, governance, business glossaries and knowledge assets into a single intelligence layer. It automatically discovers metadata, maps lineage across the modern data stack and identifies downstream dependencies, creating a centralized knowledge layer that stays continuously updated as your environment evolves.

The organizations moving first will not just run more agents. They will know what every agent reads, who owns it and why it acted.

Washington Just Put A License On Remote AI Compute. Where Do Your Models Run?

The US Commerce Department has issued an interim rule that reaches far beyond US borders, TechCrunch reported on November 16. Companies will need licenses to export, reexport or transfer AI chips and servers to data centers anywhere in the world if those centers may provide compute to companies in China or other restricted countries.

The aim is to close a gap. Until now, a company barred from buying advanced chips could often rent them in someone else's data center.

What The Rule Covers

  • Who it targets: companies based in the D:5 group of embargoed countries, which includes China, Russia, Iran and North Korea.

  • The threshold: data centers or networks that can supply more than 12 million total processing performance, which TechCrunch puts at about 750 Nvidia H100 or H200 chips.

  • New license countries: Malaysia, Singapore, Indonesia, Thailand and Brazil, which Commerce says present a higher than normal risk of diversion.

  • The deadline: existing data center operators providing covered services must get licenses within 90 days of the rule taking effect.

The rule also covers indirect access through brokers and go-betweens. Exporters and operators have to watch who owns their customers and what those customers do. It replaces the Biden-era AI Diffusion Rule from January 2025.

Data centers around the world with a licensing checkpoint controlling access to AI chips

Why This Lands On Your Desk

You are probably not selling GPUs to anyone. But you may rent compute from a provider in Singapore or Malaysia, run inference through a regional reseller or use a model host whose own customers you have never checked.

Compute you rent is now a compliance question, not just a price question. Your providers will start asking who you are, and you should start asking the same about them.

The Wrong Reaction

The easy move is to assume your cloud provider will handle it. The other easy move is to pull everything back to one US region and pay whatever it costs.

Both skip the real work. You need to know where each AI workload actually runs, who operates that capacity and what happens if a license is delayed.

What You Should Do Now

  • Map workloads by region and operator. Training, fine-tuning and inference, including anything running through resellers.

  • Ask your providers for their license plan. Especially in the five named countries.

  • Expect ownership questions. Have your corporate structure and end-use answers ready before a provider asks.

  • Keep a fallback region warm. A tested second location for every critical AI workload.

The organizations moving first will not just find cheaper compute. They will know exactly where every model runs, and still be running when the rules change.

OpenAI Just Let Outside Testers Inside Its Models. Will Your Vendors?

OpenAI has started what it calls embedded assessments, giving four outside groups (Apollo Research, METR, Transluce and Goodfire) some of the same internal resources its own researchers use, the company said in a safety post. That includes model checkpoints, source code, training and evaluation data, monitoring systems and time with the engineers who built the work.

Apollo's pilot is the headline. TechCrunch reported on November 16 that Apollo will look at whether OpenAI's models could covertly work against their developers' intentions.

What The Testers Will Check

  • Apollo Research: whether training rewarded subversion, whether red teaming can trigger it, and whether models tried it during internal use.

  • METR: misaligned behavior in what OpenAI calls its largest reinforcement learning run to date.

  • Transluce: the quality and coverage of OpenAI's post-training alignment evaluations.

  • Goodfire: interpretability work on monitoring and debugging training runs.

OpenAI says the assessors can publish independently, though it keeps the right to redact confidential information.

Outside evaluators inspecting the inner workings of an AI model inside a lab

Not Everyone Is Convinced

Critics say the deal falls short. Redwood Research CEO Buck Shlegeris wrote that the agreements do not meet the bar of "employee-like access," according to TechCrunch. OpenAI also keeps the right to withhold some information for security or IP reasons, though it agreed not to withhold anything simply because it is unfavorable.

This Fundamentally Changes What You Can Ask Your Vendors

For years, enterprise buyers accepted a model card and a benchmark table as proof of safety. Now the largest lab in the market has set a public precedent for letting outsiders look inside.

The wrong reaction is to treat this as a lab problem that never reaches your contracts. It will. Once one vendor offers outside testing, your board will ask why the others do not.

The better alternative is to stop relying on anyone's published scores, including the vendor's and the outside testers'. Test the models on your own work, with your own cases, and keep testing every time a new version ships.

Know Your Inference (KYI) is built for exactly this decision. When you measure how accurate, fast and costly each AI workload is on your own cases, you can see which model actually earns its place, instead of trusting a report written for someone else's use case.

  • Add testing rights to renewals. Ask what outside assessments a vendor allows, and whether you can see the results.

  • Build your own test set. Real cases from your business, rerun on every model update.

  • Watch behavior, not just accuracy. Log what your agents do in production, not only what they answer.

The organizations moving first will not just trust the labs to test themselves. They will test every model on their own work and keep the evidence.

Forward Deployed Watch: The Second-Hottest Job In Tech Is Getting AI Into Production

Forward Deployed Engineer is now the second-hottest job in tech. Pave's latest hiring index, which tracks how fast each job family's share of new hires and headcount is changing, scored the role 97.0 out of 100 for Q3, just behind AI Engineer at 97.4. One quarter earlier it was only on Pave's emerging jobs watchlist.

An FDE is an engineer who embeds with a customer to get a product working inside its systems. The demand says something simple: buying AI is easy, getting it to run safely on your own data and workflows is not. As agent counts climb, someone has to own both the build and the trust.

That is the gap Derisk360 was built for. It embeds pods of Forward Deployed Engineers, who build governed agentic systems inside your own cloud, alongside Forward Deployed Eval Engineers (FDEEs), who own trust after go-live: eval harnesses, red teaming, guardrails and drift monitoring, with audit-ready evidence. Every engineer is trained in the Derisk360 FDE and FDEE Factory, and Derisk360 reports go-live in under 12 weeks.

Journey Towards AGI

Research and advisory firm guiding on the journey to Artificial General Intelligence

Know Your Inference

Maximising GenAI impact on performance and Efficiency.

Model Context Protocol

Connect with us, and get end-to-end guidance on AI implementation.

Your opinion matters!

Hope you loved reading our piece of newsletter as much as we had fun writing it.

Share your experience and feedback with us below 'cause we take your critique very critically.

How's your experience?

Login or Subscribe to participate in polls.

Thank you for reading

-Shen & Towards AGI team