AI Dev News Digest - Sept 26th, 2025

Here's what mattered this week for AI developers: new practical tools emerged (Copilot CLI, Responses API), agent workflows became more streamlined (Chrome DevTools MCP, Data Commons MCP), and significant funding was directed toward developer-first agent platforms. For deeper insights, check out the longer reads on "context engineering" at the end.

With the new tools introduced this week, your AI engineering tasks should start feeling less like a juggling act. GitHub Copilot CLI now includes MCP support. Chrome DevTools has its own MCP, allowing you to control a real browser from your IDE using AI. You can trace, profile, take screenshots, fill out forms, and execute tasks with precision. Google's Data Commons MCP Server also provides structured public data effortlessly. These aren't just updates—they represent a quiet revolution where agents perform tasks rather than just discuss them.

Dev Tools & Agent Platforms

• GitHub Copilot CLI public preview. A terminal-native Copilot that can plan, edit, and run tasks with MCP support; you review every action before it runs. Install via npm i -g @github/copilot. (GitHub Changelog)
• OpenAI Responses API for structured output. One API for text, tools, and JSON mode with schema enforcement and re-runs; less glue code for function calling. (OpenAI Dev)
• Chrome DevTools MCP for agents. Lets your coding agent drive a real Chrome instance via MCP + Puppeteer for tracing, screenshots, network analysis, and reliable waits. (Chrome Developers)
• LlamaIndex adds document understanding to Claude Code. Brings retrieval and doc parsing utilities into Claude Code so agents can summarize, ground, and reason over project docs. (LlamaIndex Blog)
• Google Data Commons MCP Server. A first-party MCP server exposing structured public data (Data Commons) to your agent; query entities and properties without custom scrapers. (Google Developers Blog)
• NVIDIA: build a RAG agent with Nemotron. Step-by-step tutorial to assemble a retrieval agent around Nemotron models, with prompt and indexing patterns. (NVIDIA Developer Blog)
• Clarifai “reasoning engine.” Vendor says it speeds up workflows and reduces costs; worth a look if you’re chasing latency/price trade-offs. (Cosmico)
• Microsoft reimagines Marketplace for AI apps and agents. Microsoft is unifying AppSource and Azure Marketplace and adding a new “AI apps and agents” category (3,000+ listings), with co-sell and distributor integrations plus simpler publishing for agent builders. Good news for devs shipping agents inside Microsoft 365 Copilot and Azure AI Foundry. (Microsoft Partner Blog)

Models, APIs & Benchmarks

• GPT-5 Codex in GitHub Copilot (public preview). New code-focused model rolling out in Copilot; expect better planning and tool use in agent mode. (GitHub Changelog)
• NVIDIA open-sources Nemotron models and datasets. Foundation models + training data available for tuning and evals; clearer path to custom coding agents without license headaches. (NVIDIA Blog)
• Hugging Face GAIA2 + ARE 1.0. New agent benchmark and a configurable research environment for read-write, noisy, time-bounded tasks; includes traces and cost/speed metrics. (Hugging Face)
• Smol2Operator for GUI agents. Recipe and datasets to train VLMs that click, type, and navigate UIs; open source and reproducible. (Hugging Face)

Code Search & IDE

• Copilot’s new embedding model. Better semantic code search in VS Code grounded in a fresh embedder; helps find symbols and intent across repos. (GitHub Blog)
• Gartner MQ: GitHub a Leader (again). Recognition matters to buyers; for teams standardizing on Copilot, this lowers procurement friction. (GitHub Blog)

Cloud, Enterprise & Infra

• OpenAI + Oracle add five Stargate sites. More AI data centers planned under the Stargate program; hints at long-term capacity planning. (OpenAI)
• Local reporting on the $165B campus. Regional outlet details scope and siting around El Paso/Doña Ana for the OpenAI/Oracle complex. (El Paso Matters)
• Alibaba Cloud to offer NVIDIA “Physical AI” dev tools. NVIDIA’s robotics/embodied stack shows up in Alibaba’s platform, broadening access in APAC. (TechCrunch)
• Claude now in Microsoft 365 Copilot. Anthropic models available in 365 Copilot for enterprises that prefer Claude’s behavior or policies. (Anthropic)
• Komprise adds data prep for AI. New features to inventory and deliver the right unstructured data to model pipelines. (Komprise Blog)
• DESILO’s THOR framework. A privacy-safe LLM inference stack pitched for regulated workloads in Korea; keep an eye on technical details as they emerge. (Korea Herald)

Funding & Company News

• Factory raises $50M Series B. Agent-native “Droids” for software engineering, now touting top marks on Terminal Bench and multi-interface support (CLI/IDE/Slack/Linear/browser). (Factory.ai)
• Modular raises $250M. More fuel for its unified compute layer (Mojo, Max, inference runtime) aimed at squeezing perf across CPUs/GPUs. (Modular Blog)
• Flox raises $25M Series B. Nix-powered dev environments with policy and team workflows; push-button, reproducible envs for AI stacks. (Flox Blog)
• Obot AI raises $35M seed. Building an enterprise MCP gateway/infrastructure layer so agents can safely reach internal systems. (Obot.ai)
• Emergent raises $23M. “Vibe coding” platform funding to pursue natural-language app building; early but developer-adjacent. (X)
• Requesty raises $3M seed. Tooling to help developers build on LLMs with safer defaults and abstractions. (LinkedIn)

Events & Programs

• AI Engineer Paris kicks off. Workshops and talks centered on practical agent and tooling workflows. (AI Engineer Paris)
• Google Cloud Summit Switzerland + Hack Challenge. Regional summit with a hands-on challenge for builders. (Google Cloud OnAir)
• CData Connect AI Hackathon 2025. Data connectivity + AI use-cases; handy if your agent needs to talk to lots of sources. (CData Blog)
• OpenAI Grove application deadline hit. Developer mentorship program closed; watch for the next cohort. (OpenAI)
• NextCorps AI Startup Jam (Oct 24–26). Weekend sprint to build a software prototype with AI tools—no coding required; limited to 75 participants; $25 if accepted; Western NY residency required; cash prizes. Handy community event if you’re nearby and want to validate an idea fast. (NextCorps)

Legal & Policy

• xAI sues OpenAI (again). New suit alleges trade secret theft; nothing to act on now, but track any discovery that implicates training data or model provenance. (Decrypt)
• Anthropic updates developer-relevant terms. Legal blog breakdown of changes to Claude’s terms and privacy posture; skim before enterprise rollout. (AMST Legal)

Weekend Reading

• Demand for junior developers softens as AI takes over. CIO reports that AI coding assistants and low-code tools are shrinking demand for entry-level roles, with juniors expected to shift toward oversight and integration work. Worth a skim if you mentor juniors or plan 2026 hiring. (CIO)

• The Context Revolution. Plain-English look at why “context engineering” now matters more than prompt tricks. (Medium)
• Context engineering beyond prompt engineering. Forbes Council piece arguing for system-level context design in production apps. (Forbes)