The Agent Stack Around the Model Gets Its Own Launches
Endor Labs joined Google Antigravity for secure coding, Baseten partnered with Postman on agent inference, Google shipped a generative-interface standard, and Daytona argued agents need their own runtime.
- Endor Labs joined the Google Antigravity Plugin marketplace, with AURI Agents reported 2.8x faster on security tasks and using 12x fewer tokens than general-purpose agents without Endor context. Self-reported m:…0824
- Baseten is the recommended inference partner for Postman's enterprise agents, pooling GPU capacity across clouds. Self-reported m:…4786
- Google's A2UI is a framework-agnostic open standard for generative UI, describing an interface as data for runtime rendering. m:…5239
- Daytona provisions Firecracker microVMs in less than 90 ms, against Kubernetes setups that Daytona says need custom controllers for agent lifecycles. Self-reported m:…4999
Security inside the coding agent
Endor Labs is a launch partner for the Plugin marketplace in Google Antigravity, bringing its AURI Agents and its MCP server into that coding environment. m:…0824
Google was already a launch partner for Endor Labs Agent Governance in May 2026, according to Endor, so the Antigravity marketplace listing extends an existing relationship rather than starting one from scratch. m:…0824
The company says AURI Agents complete security tasks 2.8 times faster than general-purpose agents without Endor Labs context, and use 12 times fewer tokens, and its MCP server can scan projects for security flaws and exposed secrets during code generation. m:…0824
The MCP server also allows agents to check dependencies for vulnerabilities before installation, and agents run on user infrastructure using their own language-model keys and source-control credentials, with most agents read-only by default. m:…0824
Endor's agents also include software-composition remediation agents that can open pull requests with evidence attached, triage agents that confirm true positives and generate patches, a helper that gives guidance on package installation, and an explainer that translates vulnerability identifiers into plain-language remediation advice. m:…0824
Endor puts the need in stark terms, saying only about 10 percent of AI-generated code is both functionally correct and secure, which is the gap its in-workflow scanning is aimed at closing. m:…0824
Inference and interfaces for agents
Baseten is the recommended inference partner for Postman's enterprise agents, providing multi-cloud capacity management to pool graphics-processor capacity across clouds, plus training infrastructure to post-train open models for specific tasks. m:…4786
Postman's platform is used by more than 40 million developers and by 98 percent of the Fortune 500, according to the announcement, which makes the partnership a distribution route for Baseten's inference as much as a technical one. m:…4786
That scale is the point of the partnership from Baseten's side: becoming the recommended inference partner puts its capacity pooling and post-training infrastructure in front of the developer audience that already builds with Postman, rather than requiring those developers to find a separate inference vendor. m:…4786
Postman's Agent Stack includes Passport, Fabric, Astropods and Fern, Baseten has zero data retention by default, and Baseten cites Harvey post-training a 27-billion-parameter open model to match closed-source frontier performance on legal benchmarks as an example of its training infrastructure in use. m:…4786
Baseten also says its model APIs ranked among the top two providers on Artificial Analysis for throughput and latency in 2026, a claim buyers can weigh alongside the partnership itself. m:…4786
Google launched A2UI, a framework-agnostic open standard for generative interfaces, which describes an interface as data for application rendering so agents can build interfaces at runtime based on what a request needs. m:…5239
CopilotKit provides the agent-integration stack for building user-facing agentic apps, and it supports integration with LangChain, among other frameworks named in the announcement. m:…5239
Google frames the need plainly: agent-user interactions are evolving from text exchanges to rich interfaces, and CopilotKit provides a toolkit including a chat surface and runtime for A2UI, with support for React and Angular rendering and custom renderers for Vue and Slack. m:…5239
CopilotKit also supports integration with AWS Strands and Google ADK in addition to LangChain, so the standard is pitched as portable across agent frameworks rather than tied to one stack. m:…5239
A runtime built for short-lived work
Daytona argues that Kubernetes is designed for long-running microservices where instances are interchangeable, while AI agents require short-lived, task-specific environments that adapt in real time. m:…4999
Running agents on Kubernetes requires teams to build custom controllers for lifecycle management and provisioning, according to Daytona, which provisions Firecracker microVMs in less than 90 milliseconds. m:…4999
Daytona cites customer examples including AfterQuery, which reduced sandbox creation times by 86 percent, and Trajectory, which saves more than 15 hours each month by using Daytona instead of stitching together Kubernetes solutions. m:…4999
Daytona cites further examples: the Laude Institute increased experiment throughput by 100 times using its stateful environments, Athena Intelligence consolidated six specialized sandbox environments into a single runtime layer, and WolfBench minimizes idle compute costs by scheduling sandboxes across cloud infrastructure, virtual machines and bare-metal servers. m:…4999
Daytona also argues against the alternatives on their own terms, saying Kata Containers increases baseline resource consumption compared with other isolation methods, and that gVisor creates potential escape paths through exploitable interface boundaries and kernel vulnerabilities. m:…4999
Put together, the four launches cover the unglamorous middle of agent development: keeping code safe, keeping inference available and affordable, giving agents a way to show their work, and giving them somewhere safe to run. m:…0824 m:…4786 m:…5239 m:…4999
What is still unresolved
- Endor's speed and token figures are company comparisons against general-purpose agents, with no independent test in the cited material.
- No pricing is given for Baseten's inference partnership or for Daytona in the cited material.
- How widely A2UI is adopted beyond the named frameworks is not shown in the cited material.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.