Welcome to Runtime! Today: Datadog thinks IT ops professionals will prefer an AI model designed for them at a fraction of the cost of the frontier, Perplexity launches into SPACE, 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
Scout's honor: The operations engineers keeping the internet economy afloat for decades have always wanted to automate as much of that work as possible, which brought the world ideas like infrastructure as code and orchestration tools such as Kubernetes. These days agents are the talk of the tech world, and while — as usual — software developers and coding agents have grabbed most of the spotlight, there are plenty of agents designed for what Google once dubbed "site reliability engineering."
However, just like coding agents, many of those SRE agents rely on costly frontier models. This week Datadog introduced a custom model called Scout — built around Z.ai's GLM-5.3 Flash open-weight model — that it promised would outperform the frontier models when it comes to troubleshooting production issues within a much smaller budget.
"We trained Scout to reduce inference cost so that investigations are cheap enough to run an order of magnitude more often and earlier in an incident’s lifecycle," Datadog said in a blog post. It was designed to work with the company's Bits Investigations agent, and is currently available through a private preview.
"Scout’s efficiency makes it practical to extend support to more alerts and investigate complex issues in greater depth, enabling greater breadth of coverage, reducing the burden of on-call work, and freeing more time for development," Datadog said. The goal is to allow Bits to respond to incidents faster and more decisively than a human SRE could, but will probably need a lot more testing before most companies are ready to hand over sole responsibility for incident management to an AI agent.
The Stack Summit*: We're convening in London on November 4-5, for a series of exclusive workshops and fireside conversations on the rise of BYOC as a favoured enterprise SaaS deployment model; how many CDOs are consolidating their data estates with Apache Iceberg; how CISOs at FTSE 100 scale are handling supply chain risk (with GSK's CISO) and more.*
Ticket applications are subject to pre-vetting. Get in touch with ed@thestack.technology if you want to be in the room.
Delivery, continued
Castles in the sandbox: If enterprises are going to ever trust the use of personal agents that connect to corporate data, they'll need to make sure that data stays within the lines. Sandboxes are the preferred method for agent containment, but after OpenAI showed the world what can go wrong with a faulty sandbox this summer, Perplexity released the technology behind its own SPACE sandbox design as an open-source project.
"Open-sourcing SPACE gives developers a production-tested foundation for agent runtimes, with isolation, state recovery, and credential and network controls built in," Perplexity said in a blog post. Available under the Apache 2.0 license, SPACE (Sandboxed Platform for Agentic Code Execution) builds on top of the open-source Firecracker microVMs designed by AWS nearly a decade ago.
Harness racing: AI harnesses have emerged as a crucial layer in the enterprise AI stack to handle the interaction between the AI model and business data, but thanks to the closed-door policy enacted by Anthropic and OpenAI around their most powerful cybersecurity models, most companies that want to use those models for defensive work are forced to pay a vendor. This week Chainguard got in on the action.
Chainguard Harness is "a new cyber harness that uses frontier AI models like Anthropic's Claude Mythos 5.1 to find and fix exploitable vulnerabilities in an organization's application code," Chainguard said in a press release. It's available through the Claude Marketplace, which means companies can use their Anthropic budgets to purchase the harness, and the band played on.
Tiny bubbles: Even small businesses like scrappy enterprise tech media publishers generate a lot of documents for one-off projects or internal presentations that start to pile up. But TinyLittleOffice.com, the latest idea to spring out of Paul Ford's brain, is a method for turning documents into links, and links into documents.
The idea is that the data within the document is stored within a link that calls a server to render that data in your browser; nothing is actually stored on the server. Editing the document replaces the old link with a new link, and while it doesn't allow for the kind of collaboration that has made tools like Google Docs so widely used, it could be a lightweight (and private) way to pass around ideas within a team, division, or organization.
The rest of The Stack
Supporting role: Third-party SaaS support providers like Rimini Street and Spinnaker Support enjoyed a growth surge during the pandemic cloud boom, but that growth has slowed in recent years amid a broader slowdown in SaaS and fears of AI disruption. Still, they're putting on a brave face as newer concepts like harness maintenance and headless software present opportunities to help out latecomers to the AI boom.
Agent takeover: At the beginning of 2026, coding agent activity on Vercel's development platform made up about 3% of overall traffic. At this point, it now accounts for 60%, and the company is evolving its product strategy in response, founder and CEO Guillermo Rauch told The Stack.
Point break: Identity management services are going through profound change as agent activity reaches the heights described by Rauch, and will become a huge security concern as the sector evolves. Proofpoint shelled out $6.3 billion this week for Varonis Systems in order to improve its approach to verifying agent activity on customer networks.
Quote of the week
"The biggest exposure is the gap between monitoring and legal accountability: organizations may be unable to detect or demonstrate how an AI system caused a harmful outcome. Even if legal consequences are shared, the deploying enterprise will usually carry more immediate reputational risk because it selects the use case and controls the customer-facing interface." — Thomas Randall, research director at Info-Tech Research Group, explaining to The Stack how the cloud-era shared responsibility model might evolve in the enterprise AI era.
We're also reading:
Claude reveals ECDSA curve origins: If you were worried that the NSA had a backdoor into a popular encryption standard, Anthropic tried to put those worries to rest.
Nowhere to hide: Tech selloff hits software and chip stocks alike: Money ain't got no owners, only spenders.
Thanks for reading — see you Tuesday!