Microsoft's agent platform, Argentina's network state, and Anthropic's AI pause

Every year at their annual developer conference, Microsoft announces a huge array of new features and services, but as the year progresses very little seems to have a major impact. It remains to be seen if this year will be any different. The announcements this week were certainly ambitious, but their track record makes us healthily sceptical.

The most significant software announcement has some promise. Microsoft has opened up "Work IQ" APIs, making the intelligence layer behind Microsoft 365 available to any external developer building agents. It means third-party agents can now tap into the kind of organisational memory that made Copilot attractive in theory. There will hopefully be more to come, our initial tests saw a few gaps and incompatibilities. Let's hope they stay the course and eventually offer the rich context layer that they are so well positioned to deliver.

“With Work IQ, we have exposed what is perhaps the most important database in a company that never got used as a database because it was only captive to our apps.”

Scout is the first of a new category Microsoft is calling Autopilots: always-on background agents built on the OpenClaw framework, with their own persistent identity and memory. Scout connects to Teams, Outlook, OneDrive, and SharePoint and is designed to handle preparation work, meeting briefings, expense reports, follow-ups, without being explicitly invoked. It is currently in preview for Frontier programme subscribers only, which means anyone not already paying for GitHub Copilot at the higher tier will not see it for some time.

Microsoft also announced seven new MAI models, including MAI-Thinking-1, its first in-house reasoning model. At 35 billion active parameters (around a trillion in total, in a sparse mixture-of-experts design) with a 256k token context window, it targets multi-step reasoning and code generation, and Microsoft claims it scores among the strongest in its class. The broader MAI family covers image generation, voice, transcription, and coding. This is Microsoft reducing its dependency on OpenAI, which has been a clear strategic direction since late 2024 but is now being formalised.

Microsoft's open-source announcements were probably the most immediately interesting. ASSERT, short for Adaptive Spec-driven Scoring for Evaluation and Regression Testing takes plain-text descriptions of how an agent should behave and turns them into executable test suites automatically. The four-stage pipeline, covering systematisation, taxonomisation, test generation, and inference scoring, produces scored results with policy citations and failure rationales, without requiring a developer to write individual test cases by hand. The Agent Governance Toolkit (AGT) is the runtime enforcement layer for controlling agents. It is a seven-package middleware system that intercepts every tool call, resource access, and inter-agent message before it executes, evaluating it against a policy engine in under 0.1 milliseconds. Policies are defined in YAML, OPA Rego, or Cedar, making them versionable, reviewable, and auditable by compliance teams rather than buried in prompt instructions.

CEO Satya Nadella's framing is not that Microsoft has the best models or the most popular products. It is that every company should be able to build its own frontier intelligence, with its own private evals, its own traces, its own harness, and its own compounding advantage. The goal for Microsoft is to deliver a platform that makes that possible for any organisation, regardless of starting point. On the No Priors podcast Nadella admitted that the industry, including Microsoft, badly underestimated the real-world complexity of deploying AI in ways that actually deliver business value.

But Microsoft's stock has declined while competitors have risen. Copilot adoption remains weak. Anthropic has taken the lead in coding tools and Microsoft's response, pulling Claude Code licences and pushing developers toward Copilot, was not warmly received. GitHub, one of Microsoft's most important developer assets, has suffered serious outages, with bot traffic overwhelming infrastructure in ways that have pushed some long-standing users to reconsider their commitment.

Whether the execution matches the architecture is a separate matter, and Build has a long history of answering that question badly.

Takeaways: The most useful things Microsoft announced at Build 2026 are not the headline products. They are the components. Enterprises building agent systems do not want to be locked into a single vendor's stack, and the platform that wins will be the one that provides the interoperability layer, the identity controls, the policy enforcement, and the eval infrastructure. Copilot Cowork + Scout point in an interesting direction, with genuine ambition around always-on, background intelligence that learns how a specific organisation works. But both are restricted in access, incomplete in capability, and untested at production scale. Microsoft is moving faster to translate concepts into something resembling a coherent system than it has in recent years. Whether they can execute is a question that the Build conference cannot answer.

Argentina's Congress is weighing a bill that would make it the first country in the world to let companies run with no humans inside them. Meanwhile the American billionaire Peter Thiel has been spending considerable time in the country with the President, Javier Milei, and has recently moved his family to Buenos Aires.

The bill creates a new legal category: the non-human corporation. These are companies run entirely by AI agents or robots, with no human owners, directors, or shareholders. Under the proposed law, the software itself can own assets, sign contracts, and pay tax. No country has done this. The European Union considered electronic personhood and rejected it. Several American states ban it outright. Argentina would be first.

Milei made the pitch himself, in a Financial Times op-ed co-written with his deregulation minister, Federico Sturzenegger. He compared Buenos Aires to Amsterdam in 1602, the birthplace of the Dutch East India Company, and invited the world's AI firms to build there free from what he called the deadly hand of premature regulation. The selling points are concrete: cheap energy for data centres, lithium for batteries, low wages, and an absence of rules.

“The machine and the legal entity were, together, the double helix of modern prosperity”

No stranger to libertarian ideas, Thiel has for two decades funded the dream of "exit". He backed seasteading, then Próspera, the charter city in Honduras, then Praxis, a self-styled network state that raised $525 million. The intellectual scaffolding comes from a 1997 book, The Sovereign Individual, which predicted that technology would let the wealthy slip free of the nation-state. The plan was always to build new countries, or carve private zones out of weak old ones.

Argentina may be accelerating that plan. Instead of seceding, Thiel may be thinking of converting an entire country into a free zone. Milei has floated replacing civil servants with AI, signed deals to run state services on foreign cloud infrastructure, and built an AI unit to monitor social media. The charter-city model needed an island. This needs a nation, and the government of Argentina is volunteering.

But there are unresolved issues with the Milei/Thiel plan. A normal company is a legal person you can sue, but real people are formally connected to it: directors who answer, owners who can be taxed or reached in court. A humanless company removes them. When one of these entities causes harm, there is no one behind the veil. Argentine academics point out that the country has no comprehensive framework for data protection or liability to catch what falls through. The law answers the question of who profits. It does not answer the question of who pays.

Takeaways: For thirty years the tech right wanted to escape the state. Argentina shows the next move is to capture one instead. A struggling country is selling the one asset rich nations refuse to offer, the absence of rules. Argentina is now a sandbox for the hardest questions in AI. Whether it counts as a person and who carries the blame? A fascinating experiment, but sadly, the people conducting it have every reason to want the loosest answer, not the safest one.

This week's chart tracks lines of code merged per engineer, per quarter, at Anthropic. From 2021 through to early 2025, the bars sit flat, barely moving. Then they go near-vertical. By Q2 2026, Anthropic's engineers are merging eight times as much code per quarter as they did before 2025. More than 80% of it was written by Claude.

In May, when we covered Jack Clark's "perspiration principle": his argument that the routine, iterative work of AI research was becoming tractable enough for machines to take on. Clark, who co-authored Anthropic's paper published this week, put the probability of recursive self-improvement at 60% by 2028. Looking at that chart, we wonder whether even that estimate is now too conservative.

Anthropic is simultaneously publishing this data and calling for a verifiable global pause mechanism, a kind of arms-control treaty for frontier AI development. The tension in that position is the point. They can see, from their own internal numbers, exactly how fast this is moving. Detractors have countered that this is more pre-IPO theatre. Only time will tell.