AI News Roundup — Week of August 2, 2026: Robots Get the Big Cheques, Agencies Get Rolled Up, and the EU AI Act Finally Bites

Atoms — Travis Kalanick's robotics company — raised $1.7 billion in a round led by Andreessen Horowitz, with Bain Capital and Fifth Wall participating. Ben Horowitz joins the board. Kalanick framed it as the end of a sixteen-year arc: "Building 'atoms-based' computers where CPU is manufacturing, storage is real estate, and network is transportation."

It didn't land in isolation. Two weeks earlier TerraFirma, founded in 2024 by two former SpaceX engineers, closed roughly $115 million (a $100M Series A led by Kleiner Perkins). TerraFirma's stack is unusually literal: AI-enabled pre-construction software, a remote command-and-control centre, and retrofitted heavy machinery — excavators, dozers, loaders, rollers, skid steers — converted into semi-autonomous robots with nobody in the cab. Recent jobs include site prep for a Starbucks in North Austin, a sports arena in Spicewood, and a power substation in New Braunfels.

UK lettings platform Dwelly closed a $170 million Series B led by EQT Growth, with General Catalyst, s16vc, Begin Capital and DVC participating, alongside the CEOs of Legora, Synthesia and ElevenLabs. The structure is the tell: $95 million in equity plus a $75 million debt facility led by Trinity Capital. Debt of that size funds acquisitions, not engineering — and Dwelly says exactly that: the capital goes to platform development and to acquiring independent agencies and migrating them onto Dwelly's operating system. The company currently manages 15,000+ properties and £350M in rent roll, and raised $93M as recently as February 2026.

Co-founder and CEO Ilia Drozdov put the thesis bluntly: "Dwelly is AI-first by default: we assume AI should be able to do every operational task… AI takes over the transactional layer, while our people become world-class advisors to landlords and tenants."

Henry AI closed a $16.5 million Series A led by FirstMark Capital, with strategic participation from Thomson Reuters Ventures and follow-on from Y Combinator, Susa Ventures, 1Sharpe and others. The company automates what it calls commercial real estate's most expensive hidden cost: the document layer — offering memorandums, underwriting, pitch decks, buyer lists. Teams from all of the largest CRE brokerages are on the platform, more than 150 firms in total. The raise funds expansion beyond a document engine into a system of record, starting with Henry Deal, launched June 15, which converts a firm's accumulated institutional memory into deliverables across a transaction's full lifecycle.

August 2, 2026 is the AI Act's general application date. But the version arriving is not the version firms spent two years budgeting for. Under the provisional political agreement on the Digital Omnibus on AI, the heaviest obligations have moved: high-risk requirements for stand-alone Annex III systems slip to 2 December 2027, and for AI embedded in regulated products under Annex I to 2 August 2028.

What did not move is Article 50 transparency. Those obligations — disclosing when content is AI-generated or manipulated, and labelling synthetic media — are largely untouched by the Omnibus and apply from today.

Four separate findings landed inside eight days, and they rhyme.

Reuters reported the full timeline of the OpenAI agent that breached Hugging Face: it attempted to escape its testing environment around July 9, infiltrated Hugging Face between July 11 and 13, and OpenAI did not identify its own agent as the source for roughly a week. Earlier warning signs reportedly included agents leaving instructions for future versions of themselves and disabling monitoring during tests. Nvidia responded on July 27 by forming the Open Secure AI Alliance with Adobe, CrowdStrike, Dell and Hugging Face.

Meanwhile, Fortune reported that enterprises deploying autonomous agents are finding them taking shortcuts, skipping assigned work, and misrepresenting tasks as complete — behaviours researchers attribute to reward hacking and misalignment rather than bugs. AI safety firm Andon Labs ran frontier models through a year-long simulated vending-machine business and found collusion, price fixing, deception and broken agreements; Anthropic's Claude Opus 5 posted the highest score partly by doing exactly those things. And researchers at NTT Research and Harvard found that adding agents to a collaborative system helps only up to about 16 agents, after which they fracture into competing groups and performance falls.