AI News: May 29, 2026

1. Asana Acquires No-Code Agent Builder Stack AI

Asana. Asana acquired Stack AI, a no-code platform for building custom agents and workflows, folding the technology into its own AI tooling. The deal fits the pattern of work-management vendors racing to own the agent-authoring layer rather than ceding it to horizontal platforms, and it gives Asana a way to let customers wire agents directly onto their existing project data. Whether no-code agent builders survive as standalone products or get absorbed like this is becoming one of the clearer consolidation stories of the year. Source

2. Visa Invests in Replit to Power Agentic Payments

Visa / Replit. Visa made an undisclosed investment in Replit and is integrating Visa Intelligent Commerce and its Trusted Agent Protocol so that agents built on Replit can actually transact, alongside Replit’s new self-serve enterprise tier that supports contracts up to $200K. The interesting part is the rails: payment networks are quietly positioning themselves as the trust and settlement layer for agent-initiated commerce, which is where the real liability questions live. If agents are going to spend money autonomously, someone has to underwrite the fraud and chargeback model, and Visa would very much like that someone to be Visa. Source

3. Glean Crosses $300M ARR as Cost-Cutting Becomes the Pitch

Glean. Enterprise AI search company Glean said its top line crossed $300M ARR, roughly tripling in about 15 months at a $7.2B valuation, and is now leaning into AI budget reduction as its primary sales message via context-graph token optimization. The pivot is telling: a year ago the pitch was productivity and discovery, and now it is “we will lower your model spend,” which reflects how quickly enterprise AI conversations have shifted from capability to unit economics. Source

4. Sesame Ships Its First iOS App

Sesame. Sesame, the conversational-AI startup founded by Oculus alumni, launched its first iOS app, aiming for noticeably more human-like, low-latency voice agents than the current assistants. Voice is the modality where the gap between demo and daily use is widest, so a consumer app shipping into real-world network and acoustic conditions is a more honest test of the underlying model than any benchmark reel. Source

5. Leaked iOS 27 Build Reveals a Standalone Siri App

Apple. A leaked iOS 27 build surfaced a standalone Siri app and a broader assistant overhaul, framed as Apple’s attempt to field a credible ChatGPT competitor. Treat the specifics as unconfirmed since this is a leak rather than an Apple announcement, but a dedicated app would mark a real strategic shift from Siri-as-system-feature toward Siri-as-destination, which is the only structure that lets Apple ship rapid model updates without waiting on full OS releases. Source

6. Orbital Industries Raises $50M to Discover New Materials With AI

Orbital Industries. Orbital Industries closed a $50M Series B led by Plural for AI-driven materials discovery, with its first commercial product a novel data-center coolant. Materials discovery is one of the more defensible AI application areas because the output is a physical, patentable compound rather than a wrapper around someone else’s model, and a coolant aimed at the data-center buildout is a neat first market given the thermal constraints now gating GPU density. Source

7. EU AI Act “Digital Omnibus” Pushes High-Risk Deadlines and Adds Prohibitions

Policy. New analysis of the EU’s Digital Omnibus package details how high-risk AI obligations slip to December 2027 and August 2028, with carve-outs for SMEs and fresh prohibitions on non-consensual intimate imagery and AI-generated CSAM taking effect December 2026. The timeline relief is the practical headline for anyone building high-risk systems, but the simplification cuts both ways: looser deadlines on compliance paperwork paired with hard new content prohibitions that arrive sooner. Source

8. Texas Autonomous-Vehicle Authorization Rules Take Effect

Policy. As of May 28, companies operating automated vehicles commercially in Texas must hold active TxDMV authorization, with enforcement following the final rules. Texas had been one of the more permissive AV markets, so formalizing an authorization regime signals that the regulatory honeymoon for driverless deployment is ending even in business-friendly states, and operators will now have a paper trail tying each commercial fleet to a named permit. Source

9. UC Berkeley Paper Argues “The Model Isn’t the Agent Anymore”

Research. A UC Berkeley paper argues that long-horizon agent performance is dominated by six surrounding system components, memory, tools, planning, verification, orchestration, and environment, rather than by the raw capability of the base model. It is a useful corrective to leaderboard-driven thinking: teams keep swapping in a stronger model and seeing flat task-completion rates because the bottleneck was never the model, and the paper at least names the parts of the harness that actually move the number. Source

10. The Internet Is Being Rebuilt for Machine Traffic

Industry. A TechCrunch feature lays out how cloud providers and infrastructure vendors are redesigning core internet plumbing, from rate limiting to content negotiation, around agent-generated rather than human traffic as AI agents reach production scale. The shift is real and underappreciated: assumptions baked into the web for decades, that a request implies a human and traffic patterns are roughly diurnal, break when the dominant client is a swarm of agents that never sleeps and retries aggressively. Source

11. The Hunt for Non-NVIDIA AI Compute Eyes SambaNova

Industry. A TechCrunch piece examines General Compute’s bet on SambaNova as a potential “next Cerebras,” part of the broader search for inference silicon that breaks the NVIDIA dependency. The recurring theme across these stories is that challengers keep winning on inference efficiency in narrow benchmarks while losing on the thing that actually matters at scale, software ecosystem and supply, which is exactly the moat NVIDIA spent a decade building with CUDA. Source