OpenAI Agents API Released: Codex Harness Productized as Standalone Infrastructure
On September 10, OpenAI released the Agents API in public beta. Unlike the conventional simple model API, it exposes the agent harness used in Codex directly as an API, letting OpenAI manage context, tool calls, subagent orchestration, and long-running state. For the execution environment, you can choose not only OpenAI's hosted sandbox but also external sandboxes such as Cloudflare, E2B, Modal, and Vercel, or your own infrastructure. The key point is that the unit of abstraction for agent development is shifting from a model call to a persistent execution harness.
frameworksVS Code 1.137β1.138: Agent Moves from Chat Feature to Actual Development Runtime
On September 9, VS Code 1.137 added Agent Automations Preview, enabling agent tasks to be scheduled on an hourly, daily, or weekly basis. Then on September 16, version 1.138 made it possible to run agent sessions directly inside a Dev Container, allowing the project's actual compiler, dependencies, and toolchain to be used as-is. Codex sessions were also extended to carry over across multiple apps. This marks a shift in IDE AI capabilities from "conversational assistant" toward a continuously running development worker.
infrastructureReact 19.3: View Transition and Fragment Refs Promoted to Stable
With the release of React 19.3 on September 9, the previously experimental ViewTransition and Fragment Refs became stable APIs. `<ViewTransition>` combines React Transitions with the browser's View Transition API, enabling declarative animation of a component's mount, unmount, movement, and resize. React's direction of absorbing browser-native visual transitions into its reconciliation model, beyond just state management and rendering, has become increasingly clear.
frameworksCUDA 13.4: Windows on Arm Support and Early Rubin Preview
On September 9, NVIDIA released the CUDA Toolkit 13.4. It formally introduced CUDA development support for Windows on Arm and includes preview support for the next-generation Rubin GPU architecture. MPS V3 gains TOML configuration, named server instances, and cgroup-based GPU memory limits, enabling fine-grained control over GPU partitioning in container environments. CUDA continues to expand from a simple kernel programming toolkit toward a heterogeneous platform and multi-tenant GPU runtime.
infrastructureNVIDIA Releases CUDA Rust: Extending Rust Ownership to GPU Kernels
On September 8, NVIDIA released two projects for writing CUDA kernels in Rust: `cuda-oxide` and `cutile-rs`. `cuda-oxide` compiles Rust kernels directly to PTX through a custom `rustc` codegen backend, while `cutile-rs` provides a tile programming approach using CUDA Tile IR from stable Rust. The central design goal is constraining aliasing and tensor partitioning through Rust's ownership model, illustrating how the domain of memory-safe systems languages is expanding from CPU system programming all the way to GPU kernel programming.
languagesMicrosoft Elevates Rust to a First-Class Systems Language in Earnest
In 2026, Microsoft's adoption of Rust has moved past the experimental stage and into genuine first-class systems language status. The Windows team has officially committed to making Rust a first-class language for Windows driver development and is building a driver toolchain that includes `cargo-wdk`. On the Azure side, the Rust SDK reached general availability in May 2026, providing stable APIs and SemVer guarantees, and the `windows` crate gives Rust direct access to the entire Win32 API surface. Microsoft Research is also rewriting the core cryptographic library SymCrypt in Rust, backed by formal verification.
Now I really need to learn Rust..
languagesRust 1.98.1 Emergency Patch: Trait Object VTable Miscompilation Fixed
On September 3, Rust 1.98.1 was released as an emergency point release. A miscompilation was discovered in 1.98.0 where, under certain conditions, a null pointer could be generated in a trait object vtable instead of a function pointer, which could lead not only to a segfault at runtime but also to undefined behavior. The incident carries significant weight as an example showing that, even in a safety-focused language, compiler correctness remains the final trust boundary, more so than any language feature addition.
general