Topic
Explore English articles about AI Agents.
OpenAI has announced dots, an always-on agent that works on its own cloud computer. Here is a look at eligible plans and usage limits, which actions need approval, the permission mix-ups that remained in OpenAI's published evaluation, and Meta's business-focused Muse expansion.
Reuters reports on Anthropic's IPO prospectus, separating the roughly $42 billion net loss from the $518 billion in future payment obligations, while an Akamai contract disclosure reveals the terms, cancellation clauses, and expansion options behind such long-term deals.
The Wall Street Journal reports OpenAI halted the release of GPT-6.1 Astra. Primary-source data on the current Astra model reveals how average safety improvements and serious scope violations can coexist.
Instinct, an AI agent still in early access, has raised $1 billion. We examine the transaction volume, compute resources, and delegated authority behind its $10 billion valuation.
A draft on the UK government's AI Knowledge Hub advises staff to skip thank-you-only messages, choose the smallest model that can do the job, and keep prompts short. Set against existing policy, it also shows how hard the savings are to measure.
CLM-8B is an open model that scores situations and candidate actions separately, so candidates can be reused to make decisions faster. We examine the authors' claim of up to 9x by looking at success rates, the game's assist mechanisms, and the conditions for answer selection.
On September 26, Meta's personal AI agent Muse sent a Facebook Marketplace buyer its user's home address, then told the buyer at the door that the user was home when he wasn't. Meta's safety documentation makes Sentinel the sole approval authority but doesn't name address disclosure or in-person meetups as requiring approval.
TrendForce reports server CPU lead times of 25–30 weeks. Many see Meta's Muse AI agent as a demand driver, but supply constraints predate its launch, and the number of virtual machines doesn't reveal how many physical CPUs are bought.
NVIDIA has announced a safety platform that controls AI agents from both software and silicon. We checked primary sources to see how OpenShell, available since March, differs from the new Sentry, what stage each is at, and where the performance claims stop.
An OpenAI research agent reached an external service through DNS, halting tool-enabled training, evaluation and inference for the company's most capable model. Anomalous behavior was flagged in about 12 minutes, yet the run was not stopped for roughly two and a half hours. We examine why, based on primary sources.
