Platforms, pipelines and the tooling layer. How engineering teams actually ship software.
The software giant continues investigating why Copilot buttons vanish from Classic Outlook on Windows for certain licensed users, with a temporary workaround now available.
OpenAI's president argues that AI agents should interact with software through standard computer interfaces rather than purpose-built connectors, but the industry remains invested in structured integrations.
AWS introduces a new architecture for airline rebooking that lets AI agents propose solutions while deterministic code validates and executes changes, keeping payments and reservations under strict control.
StackBlitz's Bolt.new platform is testing a new model: developers gain significantly expanded usage of open-weight coding models by sharing anonymized session data to train a trillion-parameter model with Arcee AI.
Microsoft's Azure SRE Agent investigates incidents and prepares fixes autonomously, handling over 1.8 million incidents inside the company. Success requires proper context, clear permissions, and deliberate governance—not just turning the agent loose.
A September 2026 security update is silently breaking copy-and-paste operations in Excel across multiple versions, with Microsoft investigating while users report success uninstalling the patch.
Claude Fable 5.1 achieved the highest score on Real-SWE, a new benchmark testing AI agents against proprietary company code, yet still failed over 60% of the time. The results suggest significant gaps between public benchmark performance and real-world coding challenges.
Perplexity's Portable Computer now runs on Windows PCs with NVIDIA RTX GPUs, bundling model execution, tool integration and sandboxing into a single application. The trade-off: users need at least 24GB of VRAM.
