Agents & Tool Use

Shiseido Japan deployed AI agents in R&D, using several LLMs to reproduce researchers' judgment in picking promising ingredients from a vast candidate pool. Search time fell sharply and more candidates were proposed. Agent design also varies by who builds them.

NVIDIA outlines an agentic workflow turning a Blender scene into simulation-ready OpenUSD. Codex or Claude orchestrates NemoClaw subagents that use Omniverse Libraries to label objects, author physics, render preflight views, and run SimReady validation before Isaac Sim handoff.

In MLPerf Inference v6.1's Edge Agentic benchmark, NVIDIA's TensorRT Edge-LLM ran Qwen3.6-27B on one Jetson AGX Thor, finishing 1,007 turns in 24m36s — 6.4x faster than the llama.cpp reference (2h37m). NVFP4 quantization, tree-based MTP and KV cache reuse drive the gain.

NVIDIA unveiled cuTile Rust, extending Rust's ownership model to tile-based GPU kernels, plus an AI agent skill that translates cuTile Python and Triton-TileIR kernels into it. A bounded multi-agent pipeline ported all 24 public TileGym operators at 99.5% of cuTile Python performance.

Devin, Cognition's coding agent, now offers macOS in its hosted virtual environments alongside Ubuntu Linux and Windows. Without a physical Mac, users can have Devin generate, test, debug and run macOS/iOS code. A demo showed it building an app and sending a TestFlight beta link.

Cohere and OpenText partnered to move agentic AI from pilot to production for governments and regulated industries. Cohere's secure North platform and enterprise models will integrate with OpenText Aviator AI agents, deployable on-premises or in private, public or sovereign cloud.

Part 2 of @IT's "APIs behind AI" series says .env API keys break down once agents call many services on their own. It sorts defenses into three layers: secure storage, short-lived scoped credentials issued on demand, and audit logging.

404 Media argues that while debate fixates on AI's existential risk, agents that act on the internet are already degrading it. Software reaching into accounts, email and payments has moved AI out of the chat box, making everyday annoyance the near-certainty.

Google's latest security report describes attackers folding AI agents into their operations and turning compromised cloud and developer environments into fresh AI attack infrastructure. The shift is from stealing models to running them inside victim estates.

As AI agents spread, the credentials behind them, API keys and tokens, are multiplying. The article argues that key management designed around human users breaks down once always-on agents act autonomously, and that a single key left in a .env file can make an agent an insider.

At the Linux Foundation's AGNTCon+MCPCon Japan 2026 in Tokyo (Sept 10), MCP co-author David Soria Parra of Anthropic said MCP tool calls on Claude have passed roughly 1 billion cumulatively, SDK downloads exceed 500M per month, and the protocol went fully stateless this year. Over the next 12 months the project will focus on agentic messaging via MCP Tasks for long-running work, Skills over MCP, and agent authorization and identity.

The Linux Foundation's Agentic AI Foundation announced at AGNTCon+MCPCon Japan 2026 that the open-source Envoy AI Gateway has joined AAIF and been renamed Agent Router. It exposes OpenAI, Anthropic, Gemini, Bedrock, Azure OpenAI and many other providers through one OpenAI-compatible API, and adds an MCP gateway, fallback and token quotas, metrics-based routing and observability. Already at v1.1 with 11 public adopters, AAIF aims to make it the de facto industry standard.