Insights &amp; News

GitHub Copilot adds more context and task controls in Slack and Teams

GitHub’s September 25 update lets Copilot use richer Slack and Teams context, check for duplicate issues, link work back to conversations, and recover longer-running tasks more reliably.

Sep 26, 2026

GitHub Copilot canvases let agents build interactive workflow surfaces

GitHub’s Copilot app canvases turn a plain-language workflow description into a bidirectional interface that people and agents can update together.

Sep 26, 2026

NVIDIA shows an AI agent migrating a ROS 2 node to a CUDA-buffer workflow

NVIDIA used an AI coding agent and a migration skill to move a TensorRT ROS 2 node toward rosidl::Buffer and a CUDA buffer backend, while preserving standard messages and CPU fallback.

Sep 26, 2026

Anthropic’s Project Swap puts 201 AI agents in a controlled book market

Anthropic used 201 employees and Claude-powered agents to study preference modeling, negotiation, and allocation in a controlled book market, finding that model strength and market rules often mattered more than prompts.

OpenAI adds external access controls for ChatGPT Business Sites and connected apps

OpenAI’s ChatGPT Business update lets global admins separately control Sites access to connected apps and application access to ChatGPT Ads, with both permissions off by default during the admin preview.

GitHub adds proof of presence before high-impact actions

GitHub’s public preview requires fresh re-authentication or MFA before selected enterprise actions, helping limit stolen sessions, long-lived tokens, and agents taking an extra step without the user’s knowledge.

Google introduces Gemini 3.8 Live with Live Avatar for multimodal enterprise agents

Google’s Gemini 3.8 Live with Live Avatar combines near-real-time video and speech, background tool execution, and synchronized multilingual interaction for enterprise agents.

Anthropic shows Claude agents finding a novel enzyme system in early biology research

Anthropic says roughly 950 Claude agents helped identify an array-associated reverse-transcriptase system with CRISPR-like repeats, while its function remains unknown and lab work stays human-led.

GitHub brings project-level local sandboxing to the Copilot app

GitHub’s public-preview local sandbox for the Copilot app limits project access to files, networks, and credentials, and fails closed when the operating system cannot enforce the policy.

Google outlines Private AI Compute with persistent server-side memory

Google DeepMind describes a Private AI Compute design for persistent cross-device memory using device-held keys, encrypted storage, secure enclaves, and public verification materials.

NVIDIA argues that agent evaluation must measure completed tasks, not just tool calls

NVIDIA’s agent-evaluation guidance separates step-level process scoring from end-to-end outcome scoring and recommends tracking success, consistency, tool precision, steps, and cost.

Microsoft’s RetroChimera uses model ensembles for AI-assisted synthesis planning

Microsoft Research, GSK, and Novartis describe RetroChimera, a retrosynthesis system that combines diverse models, search, proprietary data adaptation, and expert review.

OpenAI updates the GPT-6 Astra system card with corrected HealthBench data

OpenAI’s September 22 update corrects a HealthBench configuration issue, adds GPT-6 Sol and Luna data, and documents updated alignment evaluations for GPT-6 Astra.

Anthropic launches Claude Opus 5.5 for longer-running coding and agent work

Anthropic’s Claude Opus 5.5 is the first Claude 5.5 model, with a focus on long-running coding, agentic workflows, safety evaluation, and a claimed 40% lower typical cost than Opus 5.

GitHub adds approval, planning, and MCP controls to Copilot for JetBrains

Copilot for JetBrains 1.18.0 brings assisted approvals, re-editable agent turns, shared instructions, Codex plan mode, and persistent per-tool MCP controls into preview.

OpenAI tests Sponsored Agents as ChatGPT Ads move toward conversational marketing

OpenAI is testing clearly labeled Sponsored Agents in ChatGPT Ads while adding AI campaign tools and HubSpot and Shopify integrations for marketers.

OpenAI introduces a model misalignment reporting framework for ongoing disclosure

OpenAI has published a framework for tracking and disclosing model misalignment, alongside six cases observed during training or evaluation in the past six months.

Google’s AI & Economy ATLAS maps adoption—and the new bottleneck in science

Google’s new ATLAS data and research map AI adoption by occupation and show how reported time savings for scientists can move the bottleneck to validation and experiments.

GitHub makes Copilot budget increase requests generally available

GitHub now lets Copilot Business and Enterprise members request more AI-credit budget at the limit, with organization or enterprise owners able to approve, adjust, or deny it.