Welcome to Runtime! Today on Product Saturday: Google has a few ideas for the new cloud infrastructure services that agentic AI workloads will likely need to run efficiently, DuckDB turns 2.0, and more.
Please forward this email to a friend or colleague! If it was forwarded to you, sign up here to get Runtime for free every week, or level up here.
Ship it
Engine engine #9: Virtual machines and containers allowed cloud infrastructure customers to scale computing resources as their businesses grew, compared to pre-cloud strategies that involved buying a bunch of servers in hopes the people come, only to run out of capacity when too many of them show up on a Tuesday. But agents are a new beast, and they present new challenges for enterprise architects looking to wring as much efficiency out of their infrastructure spending as possible.
This week Google Cloud introduced a new infrastructure service called Agent Compute Engine, which was designed for the unique ways agents consume computing resources. "Agent Compute Engine arose from the need to handle Google’s massive agentic workloads that write and execute code, run tools, drive browsers, orchestrate enterprise workflows, and iterate rapidly and at scale," and it will soon be available to customers to use with their own agents, Google said in a blog post.
Agents operate in bursts of activity, and Agent Compute Engine takes advantage of microVMs (Firecracker is a notable example) and "copy-on-write memory sharing" to allow agents to start and stop more quickly than traditional cloud resources can accommodate. It also runs on Google's Titanium hardware security chip, providing "silicon-rooted defense-in-depth built for untrusted AI agents," which is basically a fancy sandbox.
Google thinks this architecture will also help companies run agents for longer periods of time, which requires them to have a secure, persistent connection to corporate data stores. "By weaving Agent Compute Engine directly into Google’s cloud and identity fabric, enterprise connectivity is governed by the exact same security and identity model that protects the microVM itself, enabling a unified identity model as well as auditability," Google said.
Scale AI with confidence on the infrastructure built for it. Akamai Cloud delivers high-performance cloud computing with enterprise-grade security, core to edge. Learn more.
Delivery, continued
Speaking of engines: Neo4j's graph databases have been a good fit for the AI era, helping companies organize their data and track agent activity. This week it introduced the Neo4j Context Engine, which promises to help agents wade their way through both structured and unstructured data.
"AI agents ... have the potential to massively accelerate enterprise operations through automation, but don’t inherently understand what 'customer,' 'account,' 'supplier,' 'risk,' or thousands of other relationships and concepts mean within a particular organization," the company said in a press release. Available as a private preview, Neo4j Context Engine allows companies to define those business terms in agent-friendly ways and point agents toward corporate data with that context at hand.
Make way for 2.0: AWS might have acquired DuckLabs, the company behind the popular DuckDB open-source database, but the project continues on under the MIT license. This week it reached the 2.0 milestone with the release of "Cinnamon Teal," which turns out to be a duck.
"This is our biggest feature release so far, and it ships a polished version of the Quack protocol, asynchronous I/O, a first-class VARIANT type, agent mode, triggers, an extensible SQL parser and much more, built from close to 20,000 commits since DuckDB 1.5," the project said in a blog post. It also swims faster now: "DuckDB 2.0 included optimizer and engine improvements, including faster recursive CTEs and better row-group pruning," according to a documentation web page.
Secret agents: Broadcom's sweeping changes to VMware's product and pricing strategy have seen a lot of customers looking for alternatives over the past several years, but companies that still want to manage their own infrastructure in the AI era need tools to help their developers build agents on-premises. This week VMware released Tanzu Platform 11.0, which removes the complexity of writing agents directly to infrastructure resources.
The new release focuses on a lot of the same ideas as Agent Compute Engine, allowing Tanzu developers to run their agents in secure, disposable sandboxes and giving them systems for building long-running agents as well as managing agent identity. "Platform engineers and developers can trace the entire invocation lifecycle: inspecting prompt inputs, token consumption, latency, Model Context Protocol tool calls, and retrieval context to instantly pinpoint why an agent took a specific action or where an execution stalled," VMware said in a blog post.
The rest of The Stack
Chipping in: Memory-intensive AI workloads can put quite the strain on CPUs and GPUs, which is just one reason Nvidia developed the DPU (data processing unit) to help carry that load. That type of accelerator is common in cloud environments (see Google Cloud's Titanium above) and is starting to make its way into on-premises data centers that want to run AI workloads.
Certified mess: As if enterprise administrators didn't have enough to deal with this week thanks to the PHP situation, a cascading series of errors at Indian certificate authority eMudhra forced them into a good old-fashioned yak-shaving exercise. (Look it up, kids.) Mary Branscombe unpacked the debacle for The Stack.
Missed connections: Last week's supply chain attack on ConnectWise was the work of hackers affiliated with the Chinese government, Microsoft said this week, And they used an open-source penetration-testing tool called ARTEX to grease the skids.
Quote of the week
"Based on how easily I was able to find [the flaw], and how much information was out there, and how many active reports there were of exploitation, we saw enough to trigger that part of coordinated vulnerability disclosure." — Chainguard co-founder and CEO Dan Lorenc, explaining to The Stack why he decided to break the PHP patch embargo by about six hours, a decision that did not sit well with the broader security community.
We're also reading:
Securing our frontier model training environment: OpenAI announced that it will no longer allow advanced models going through the training process to access the internet, which is a super intelligent idea.
Thanks for reading — see you Tuesday!