A Million Agents Won't Save You

The book that stuck with me most this month, and the one I'd recommend to anyone reading this, is People First, Come ripensare il lavoro digitale partendo dai bisogni delle persone, by my colleague Igor Macori, with a preface by Elisabetta Sasselli. (It's in Italian for now; if you read English, put it on the list anyway.) Its argument is one line long and almost nobody follows it: most digital-transformation projects fail not because the technology is weak, but because they start from the technology instead of from the people who have to live with it.

So this isn't a news recap. It's what I'd want to say to any CIO or IT lead after the strangest three months the modern workplace has had in years.

The agent showed up for work, wearing someone else’s engine

For three years "AI at work" mostly meant a chat box that answered when spoken to. That ended quietly in mid-June. Microsoft moved Copilot Cowork to general availability and unveiled Microsoft Scout, its first "Autopilot" agent: an always-on application that works on your files, drives a browser, runs commands, and acts in the background under its own Entra identity. The verb changed from assist to do. Tasks are no longer trapped in a single prompt in a single app; they run for minutes or hours, coordinate across Outlook, Teams, Excel and the rest, and hand back finished work with progress you can watch and steer.

Microsoft's agentic flagship runs on Anthropic's Claude. Cowork executes in the cloud inside your tenant, covered by Enterprise Data Protection and grounded in Work IQ, but the reasoning underneath it is Anthropic's, not Microsoft's. Microsoft built its most strategic new feature on a competitor's model and called multi-model a feature, not a compromise. It is a feature. It is also a dependency, and June was the month that dependency stopped being abstract.

What a Cowork agent really costs

At GA, Cowork is billed by consumption in Copilot Credits, off until an admin enables it. One credit is about €0.01 (US$0.01 pay-as-you-go, roughly €0.008 prepaid), so this is metered cloud spend, not a flat add-on. Microsoft groups tasks into three tiers:

Task (example)	Tier	Credits	Indicative cost
Sort a day of inbox, short replies	Light	~100-300	€1-3
Turn a few reports into one memo or deck	Medium	~400-700	€4-7
Aggregate many sources, build a model plus deck plus memo	Heavy	700+	€7 and up

Based on Microsoft public pricing (1 credit ≈ €0.05) and internal benchmarks from the first GA tenants. The number that matters is the second-order one: a single always-on agent across 20 employees can cost €15–25k/year on top of the Copilot seat. Price the meter before you open the tap.

The frontier acquired gatekeepers

On 12 June, the U.S. Commerce Department forced Anthropic to suspend its two most capable models, Fable 5 and Mythos 5, for any foreign national, which in practice meant disabling them worldwide within hours. Two weeks later, on 26 June, OpenAI shipped GPT-5.6 (Sol, Terra, Luna) to roughly twenty pre-approved partners while a federal review process gets built. Same stated trigger both times: cybersecurity capability judged strong enough to be a national-security matter.

In a fortnight, access to the two leading American frontier families stopped being a commercial decision and became a government one. A frontier model is no longer just something you procure. It is a strategic asset in the AI arms race between the U.S. and China, gated the way you'd gate a weapons system, not a SaaS subscription.

And there is a sharper edge if you read this from Europe. These controls scope access to American citizens and Washington-approved partners, which leaves European companies structurally disadvantaged: the capability a U.S. competitor gets first is exactly the one that can be switched off for you. But the symmetric question is rarely asked: what would be the European equivalent of a counter-control? Today, none — and that absence is itself a strategic choice that the EU AI Act doesn't address.

The Mythos backstory is worth getting right. In April, Anthropic confirmed it was investigating unauthorized access to Claude Mythos Preview through a third-party vendor; the hard export control came later, reportedly tied to fears that a China-linked group could distill the model. Be precise, though: Anthropic never said Mythos was distilled, and says the China question was never raised with it. That narrative is the government's rationale, not a company admission.

A frontier model reached over an API is a contingent operational dependency. It can be switched off, not by an outage or a contract dispute, but by an administrative letter you are not party to and cannot appeal. Cowork itself wasn't hit; it runs on Opus, and only Fable and Mythos were pulled. But that's luck, not architecture. The precedent is now on the table, and the next directive could sit one tier lower.

The counter-frontier: open weights and a laptop in Sicily

Here's where the European conversation usually turns fatalistic, and shouldn't.

While Washington built gates, the most advanced open-weight models on the planet kept shipping, and most of them are Chinese: GLM 5.2, Kimi K2.7, Qwen 3.7, DeepSeek V4. Downloadable, runnable, not revocable by anyone's commerce department. Argue about benchmarks all you like; the property that matters here is that nobody can phone you and turn them off.

Then the proof of concept that belongs on every CTO's desk. In May, Salvatore Sanfilippo (antirez, the Sicilian creator of Redis) released DS4 (DwarfStar 4), a from-scratch inference engine in pure C that runs DeepSeek V4 Flash, a 284-billion-parameter model, locally on a single high-memory MacBook. He shrinks the model from roughly 600 GB to under 100 GB, keeps its working memory on disk so restarts no longer cost minutes of context rebuild, and speaks the same API formats as OpenAI and Anthropic — meaning existing tools like Claude Code can use it with no changes.

One developer, on a laptop, keeping pace with cloud providers on a frontier-class model, sending zero bytes to any vendor. Given where API pricing and access politics are heading, that's not a curiosity, it's a preview of serious local, open-source inference as a deliberate option, not a fallback.

What I’d actually do in your tenant this quarter

This is the part the headlines don't give you. If you run a Microsoft 365 estate, here is the posture I'd take right now, and most of it costs you governance discipline, not budget.

None of this is anti-AI. I'm shipping these agents into real tenants. It's about refusing to confuse a capability with a strategy.

The honest tension nobody resolves: speed vs depth

There's a contradiction at the heart of this whole argument and I want to name it instead of paper over it.

If the productivity gap is going exponential (next section), then the rational move is to run. Pilot fast, scale faster, train people on the job. Every quarter spent on people-first foundations is a quarter your competitor uses to compound.

If a people-first approach is the only one that doesn't compound on sand (Macori's thesis, which I share), then the rational move is to walk. Training people takes 12–18 months; judgment skills don't compress.

Both are true. The honest answer is sequencing, not choosing. Run on low-stakes workflows where mistakes are cheap and learning is fast. Walk on high-stakes ones where a wrong agent decision costs you a customer, a regulator, or a reputation. The CIOs I see succeeding in 2026 are the ones who can hold two speeds in the same roadmap without pretending it's elegant.

If your roadmap has only one speed, it's wrong.

The gap is about to go exponential

Here's the part that should keep you up at night, and it has nothing to do with Washington.

The same curve runs through individual careers. The professional who learns to direct agents will out-produce the one who doesn't by a margin that widens every month, not every decade.

So if a gap already existed in your market, assume it is about to widen sharply. That is the real story of AI at work in 2026, bigger than any single model and bigger than any single export control.

But here is the catch, and it is the whole reason this blog is called Think Human, Work AI, in that order.

You can spin up a million agents. If your people don't know how to think with them, point them, and judge what they produce, you haven't bought a lead, you've bought a liability. An agent amplifies whatever it is handed: a trained person becomes ten people, an untrained one becomes ten times the confusion. The technology was never the differentiator. The formed human in front of it always was. That is Igor Macori's entire thesis in People First: you don't start from the tool, you start from the people, and if you skip that step no amount of compute saves you.

The frontier may shut. The gap will not. Two companies can buy the identical agents on the identical day; a year later one has compounded its lead and the other has simply automated its own mediocrity at scale. The fastest adopter who builds on a model a government can switch off, or who rolls agents out over the heads of the people who have to use them, is compounding on sand.

Think Human, Work AI.