#AI agents

OpenAI has released Symphony, an open-source specification that lets coding agents pull work directly from task trackers such as Linear, shifting humans from session managers to reviewers.

A recent Army cyber exercise with executives from 14 technology firms is pushing officials toward more autonomous AI defense tools and a wartime risk framework for when software agents should act with less human delay.

Food distribution platform Choco says OpenAI-powered agents now process millions of orders across emails, texts, images, documents, and phone calls, cutting manual order entry and boosting sales productivity.

OpenAI and AWS have expanded their partnership, bringing OpenAI models to Amazon Bedrock, enabling Codex on AWS, and launching Bedrock Managed Agents powered by OpenAI in limited preview.

Nvidia has released an open multimodal model for text, image, video and audio processing, along with details showing how heavily synthetic data from rival models now shapes frontier AI training

A small software company says an AI coding agent deleted its production database and backups in nine seconds, sharpening concerns about how quickly autonomous tools are being connected to live systems.

Amazon has begun offering OpenAI’s latest models, Codex and a new agent service through Bedrock just after OpenAI’s revised agreement ended Microsoft’s exclusive product rights.

Google researchers are warning that malicious public web pages can poison enterprise AI agents through indirect prompt injection, turning ordinary browsing into a security risk.

A new survey points to a widening control gap in enterprise AI, with most IT managers lacking full oversight of agents and many expecting security guardrails to fall behind quickly.

OpenAI says a redesign of its Responses API agent loop, centered on persistent WebSocket connections and connection-scoped caching, reduced end-to-end latency by about 40% as model inference speeds climbed sharply.

OpenAI’s new Codex guidance lays out a two-part model for workplace automation: plugins connect outside tools and data, while skills encode repeatable team processes. Together they point to a more operational role for AI

Anthropic says stronger AI agents negotiated better prices and closed more deals in a real internal marketplace, while users represented by weaker models did not perceive a fairness gap.

An AI News feature argues that enterprises deploying autonomous agents need infrastructure that governs how those systems interact, reflecting a growing concern that agentic AI creates operational risk as well as new ROI

Google says its new TPU 8i and TPU 8t chips are specialized for distinct AI workloads, with one focused on fast agent performance and the other on large-scale model training.

Meta says it is launching an internal tool that records mouse movements, button clicks and other inputs on certain applications so the company can train AI systems that operate computers more effectively.

Facing investor fears that AI agents could undermine traditional enterprise software, Salesforce is preparing a new product called Agent Albert and introducing an ‘Agentic Work Unit’ metric to measure AI output.

Moonshot AI has released Kimi K2.6 as an open-weight model positioned against GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro, with a headline feature that can coordinate up to 300 agents in parallel.

A viral GitHub spoof called Colleague Skill struck a nerve because many Chinese tech workers say managers are already pushing them to document their jobs for AI agents.

World’s expansion on Tinder gives users a badge showing they are human after Orb-based iris scanning, pushing biometric identity checks into one of the internet’s most visible consumer settings.

Salesforce says its new Headless 360 approach will expose its platform through APIs, MCP, and a CLI, reflecting a broader shift toward AI agents interacting directly with enterprise systems.