NVIDIA Drops Nemotron 3 Ultra for Long-Running Agents

NVIDIA Drops Nemotron 3 Ultra for Long-Running Agents

NVIDIA released Nemotron 3 Ultra, a 550B-parameter open model specifically designed for persistent AI agents [4][5]. With 55B active parameters in a Mixture-of-Experts architecture, it delivers 5x faster inference and 30% lower costs on Blackwell hardware — all while handling the complex orchestration tasks that keep agents running for hours or days.

This isn't just another language model. It's purpose-built for planning, tool use, coding, and the kind of multi-step reasoning that agents need to actually get work done [4]. The Linux Foundation OpenMDW license means open weights, open data, open recipes — a direct shot at closed models from OpenAI and Anthropic.

The timing matters. As agents move from demos to production, the infrastructure needs change. You need models that can think long-term, maintain context, and execute reliably. NVIDIA is betting that open, specialized models will win over general-purpose closed ones.

OpenAI Buys Ona to Run Codex Agents in Enterprise Clouds

OpenAI acquired Ona (formerly Gitpod) to solve the "where do agents actually run?" problem [6][7]. Instead of forcing enterprises to send code to OpenAI's servers, Ona lets Codex agents operate inside customers' own cloud environments with full security controls.

This is about persistent execution and multi-agent orchestration, not just code completion [7][8]. Agents need secure sandboxes where they can run for hours, coordinate with other agents, and access enterprise systems without exposing sensitive data. Ona provides that infrastructure layer that OpenAI was missing.

The acquisition signals OpenAI's recognition that enterprise AI isn't just about better models — it's about deployment architecture. Companies want the intelligence but need to keep control of their environments. Ona bridges that gap.

Claude Code Handles Most PRs at Anthropic

Anthropic's own Claude Code tool now handles 100% of pull requests and 80-90% of code reviews internally [9][10]. The platform features autonomous loops, persistent memory, and multi-agent coordination — shifting from manual prompting to full workflow orchestration.

This isn't just a coding assistant. It's an agentic system that understands codebases, makes edits, executes commands, and manages entire development workflows [10][11]. The "/loops" feature lets it run autonomous cycles, while persistent memory means it learns and adapts over time.

When the company building the AI uses it for nearly all their code work, that's a strong signal. Anthropic is eating their own dog food and proving that agentic coding can scale to production workloads.

EU AI Act High-Risk Rules Hit in August

The EU AI Act's high-risk system obligations become enforceable August 2, 2026 [12][13]. Companies deploying AI in biometrics, critical infrastructure, education, or employment need risk management systems, human oversight, logging, transparency measures, and CE marking by that deadline.

The compliance burden is real: conformity assessments, EU database registration, and ongoing monitoring requirements [14]. Many enterprises are scrambling to classify their systems and implement the required measures. The regulatory sandbox provisions offer some relief, but most companies are on their own.

This creates a two-speed AI market. European deployers face compliance costs and restrictions that don't apply elsewhere. The question is whether the regulatory certainty outweighs the competitive disadvantage.