AI Product Updates: New Features and Releases

Nvidia introduced a safety approach that places controls outside the agent: OpenShell restricts what it can access, while Sentry independently monitors behavior and shuts it down if it crosses the boundary.

useroot.sh and farm.palmaierc.com let users build, host and test apps using only an EVM-compatible address, including automation apps and an AI harness plugin.

B.AI added Claude Sonnet 5.5 to its API and web chat, offering faster, cheaper output, a 1M-token context window, image and PDF understanding, and five reasoning levels.

Users submit an output, dataset shard or agent trace, staked experts judge whether it meets the spec and win or lose based on their calls, and the verdict comes back with an on-chain record of who judged it and how they performed.

The platform has two parts: OpenShell runs on the CPU and sets hard limits on what an agent can attempt, while Sentry runs on separate hardware and can quarantine a misbehaving agent within milliseconds; over 100 partners have signed on.

Nvidia released an agent containment platform aimed at preventing the kind of breakout that hit OpenAI in July, with more than 100 companies already on board.

TapeAPI 1.0 issues verifiable signed receipts for OpenAI/Anthropic-compatible API and MCP calls, with local verification in Claude Code and Codex.

Quanta's Agent Lab offers paper agents, recorded market replay, hard risk limits and full decision replay for testing trading agents before going live.

OSBOOK introduces an agentic finance infrastructure layer and token launchpad, reporting nearly 5,000 agent profiles on its Square and 4.87% of supply burned.

Open-source Raven 0.2.0 puts Research, Code, Design and Oncall alongside Claude Code and Codex on one task graph with shared memory, and can rewrite its own prompts and policies.

MicroFactory released a desktop robotic work cell for electronics assembly and other tasks, pitched as being deployed by an AI agent rather than an integrator to lower the barrier to industrial robots.

Nvidia released a platform for containing rogue or misbehaving AI agents and added $150B to its share buyback program, bringing remaining authorization to $235B.