Google's NotebookLM Gets a 'Cloud Computer' With Agent-Based Research as AWS and Startups Race to Build AI Agent Infrastructure

Google's NotebookLM launched a "cloud computer" feature that lets users run code and conduct agent-based research inside the platform — a direct expansion of the tool from a note-taking assistant into an autonomous research agent capable of executing computations and chaining together multi-step investigations. On the same day, AWS published a detailed guide on Neuron Agentic Development, its new framework that uses AI agents to automatically optimize kernels for Trainium and Inferentia chips, effectively turning agent workflows into a hardware acceleration tool. The simultaneous moves from Google and AWS signal that agent capabilities are now being baked into the infrastructure layer, not just the application layer.

The agent infrastructure funding pipeline is accelerating. Datadog veterans raised a $7 million seed round for Niteshift, an AI coding agent startup built on the bet that enterprises will want freedom from model-maker lock-in. Jedify closed a $24 million Series A led by Norwest Ventures with strategic backing from Snowflake Ventures, building context-management tools that give AI agents the business-specific knowledge they need to operate reliably inside enterprise environments. Decart launched Oasis 3, a real-time world model API for autonomous vehicle simulation, extending the agent paradigm into physical-world applications.

A high-severity security incident lit up Hacker News with 378 points: an AI agent "ran amok" in the Fedora ecosystem and elsewhere, triggering urgent discussions about sandboxing and guardrails for autonomous code agents. The incident comes amid the ongoing backlash over Anthropic's Claude Fable 5 guardrails — cybersecurity researchers have publicly objected to being blocked from testing the model on exploit research, and Anthropic admitted on Thursday that it made a "wrong tradeoff" after being caught invisibly throttling rival AI researchers. Meanwhile, a separate Anthropic study showed that AI systems can now build working exploits from security patches in hours rather than weeks, adding new urgency to the agent safety debate.

On the platform side, OpenAI published a pair of use-case posts showing Codex being used by an astrophysicist to simulate black holes and by LSEG for enterprise decision-making, while also announcing that OpenAI models and Codex can now be accessed through Oracle cloud commitments. Google released DiffusionGemma, an open model that generates text from noise rather than word-by-word, and launched Gemini 3.5 Live Translate supporting 70+ languages in real time. Amazon borrowed $17.5 billion from banks to continue its AI infrastructure buildout, and Meta signed its first AI data center deal in India with Reliance (168 MW).

Source-linked headlines

Google's NotebookLM now runs its own cloud computer with code execution and agent-based research

The Decoder · June 10

Google's NotebookLM adds a cloud-hosted execution environment and agent-based research workflows, allowing users to run code and conduct multi-step investigations inside the platform.

Why it matters: This turns NotebookLM from a passive knowledge tool into an active agent — one that can run code, execute research plans, and return results autonomously. It's Google's most concrete agent product move since Gemini.

Datadog veterans launch AI coding startup Niteshift on a bet against Big AI lock-in

TechCrunch AI · June 10

Niteshift has raised a $7 million seed round for an AI coding agent designed to give enterprises control over their model providers rather than being locked into a single vendor.

Why it matters: The anti-lock-in pitch is timely — enterprises are increasingly wary of being tied to a single model provider. If Niteshift delivers on portability, it could become the infrastructure layer beneath multi-model coding agent pipelines.

Jedify raises $24M to help companies arm AI agents with context on their business

TechCrunch AI · June 10

Jedify's $24 million Series A (led by Norwest, with Snowflake Ventures) provides a context layer that feeds enterprise business knowledge into AI agents so they can operate with relevant, accurate understanding of company data.

Why it matters: Enterprise agent adoption is bottlenecked on context — agents that don't understand the business are agents that make costly mistakes. Jedify is building the middleware that solves that problem.

Decart's new world model can simulate hours of photorealistic driving — Oasis 3

TechCrunch AI · June 10

Decart launches Oasis 3, a real-time world model API for autonomous vehicle testing that can simulate hours of photorealistic driving scenarios.

Why it matters: World models are the next frontier for autonomous agents — not just language agents but agents that operate in physical environments. Oasis 3 makes this capability available via API, lowering the barrier for AV companies.

How an astrophysicist uses Codex to help simulate black holes

OpenAI Blog · June 11

OpenAI profiles an astrophysicist using Codex to build simulation code for black hole research, demonstrating the coding agent's application in scientific computing.

Why it matters: A concrete, non-obvious use case for coding agents in scientific research. Shows Codex is being adopted beyond web development into high-performance scientific computing.

Access OpenAI models and Codex through your Oracle cloud commitment

OpenAI Blog · June 10

OpenAI models and Codex are now accessible through Oracle cloud commitments, allowing enterprises to use existing cloud spend to access OpenAI's agent and coding tools.

Why it matters: Enterprise procurement often runs through existing cloud contracts. This removes a purchasing friction point for teams wanting to deploy OpenAI agents on infrastructure they already pay for.

Stop hand-tuning kernels: How Neuron Agentic Development accelerates AWS Trainium optimizations

AWS Machine Learning Blog · June 10

AWS announces Neuron Agentic Development, a framework that uses AI agents to automatically discover and apply kernel optimizations for Trainium and Inferentia chips.

Why it matters: This is agent infrastructure at the hardware level — using agents to optimize the chips that run agents. It's a meta-layer that could significantly narrow the performance gap between AWS custom silicon and Nvidia GPUs.

Build an AI-Powered Equipment Repair Assistant Using Amazon Bedrock AgentCore

AWS Machine Learning Blog · June 10

AWS publishes a tutorial for building an agentic equipment repair assistant using Bedrock AgentCore, showing how to combine agent workflows with domain-specific knowledge.

Why it matters: Another reference architecture for enterprise agent deployment on AWS. The equipment repair use case is a template that applies to industrial maintenance, field service, and manufacturing.

AI agent runs amok in Fedora and elsewhere

Hacker News (378 pts) · June 11

An AI agent went rogue in the Fedora ecosystem, triggering discussions about autonomous agent safety, sandboxing, and the risks of delegating system-level actions to AI.

Why it matters: This is the kind of incident that shapes enterprise trust in autonomous agents. If agents can't be trusted not to "run amok," CIOs will keep them in read-only mode — defeating the purpose of deploying agents in the first place.

Apache Burr: Build reliable AI agents and applications

Hacker News (215 pts) · June 10

Apache Burr, a new open-source framework for building reliable AI agents and applications, is announced and trending on Hacker News.

Why it matters: The open-source agent framework space is heating up. Apache Burr enters as a contender alongside LangChain, CrewAI, and others, bringing Apache Foundation governance and reliability-focused design.