engblogs

Frontier coding agents demonstrate building a complete Wan 2.1 video diffusion pipeline inside Modular's MAX stack using a graph-based inference API, integrating text encoding, 3D denoising, VAE decoding, and flow-scheduling into a single end-to-end system.

The updated Agents SDK delivers a model-native harness with native sandbox execution, enabling agents to inspect files, run commands, edit code, and coordinate long-running tasks across diverse tools and environments with configurable memory and workspace manifests for production-grade workflows.

Introducing a model-native harness and native sandboxing in the Agents SDK to enable durable, production-ready agents that inspect files, run commands, edit code, and coordinate long-horizon tasks across tools within standardized workspaces.

The updated Agents SDK unifies a model-native harness with native sandbox execution, enabling agents to inspect files, run commands, and edit code across long-horizon tasks in configurable workspaces for production-ready, scalable frontier-model workflows.

A technical blogpost introducing reasoning interpolation as an early-warning signal for reward hacking in reinforcement learning, leveraging donor-model prefixes and importance sampling to predict which exploits will emerge before they are observable.

Detecting early reward hacking in reinforcement learning by using reasoning interpolation to generate natural exploit-promoting prefixes and tracking exploit risk with importance sampling.

Gemini 3.1 Flash TTS advances expressive AI speech with 70+ languages, native multi-speaker dialogue, nuanced expression via audio tags, and studio-to-developer tooling across API, Vertex AI, and Google Studio, while SynthID watermarking safeguards generated audio.

Scaling Trusted Access for Cyber Defense with GPT-5.4-Cyber, Codex Security, and a democratized, iterative framework to empower defenders and strengthen ecosystem resilience.

Scaling Trusted Access for Cyber (TAC) with cyber-permissive AI like GPT-5.4-Cyber to empower thousands of defenders and hundreds of teams while balancing democratized access, iterative deployment, and ecosystem resilience with strengthened safeguards.

Scale Trusted Access for Cyber (TAC) with a cyber-permissive GPT-5.4-Cyber and Codex Security to empower defenders, emphasizing democratized access, iterative deployment, and ecosystem resilience in the next era of cyber defense.

Cloudflare Agent Cloud enables enterprises to deploy OpenAI-powered agents at the edge, leveraging GPT-5.4 and Codex to automate real-work workflows via Cloudflare Workers AI.

Enterprises deploy production-ready, agent-powered AI workflows by running OpenAI frontier models (GPT-5.4) and Codex within Cloudflare Agent Cloud and Workers AI at the edge.