OpenAI's Agents API adds command auto-review, Claude Code fences writes to the working folder, and NVIDIA's open transcription model

The Agents API is OpenAI's managed agent runtime, where OpenAI runs the Codex harness for you. Five things shipped this week. Auto Review checks shell commands before they run, the same way Codex does. Async Compaction moves conversation compaction (summarizing a long history to shrink it) into the background. Time awareness, an experimental feature, gives the model a sense of time passing. Sessions can now change their tool list mid-session, and system prompts are cached automatically across all sessions. In the Python SDK, the compaction, review and experimental settings are available from v3.42.0.

Handing an agent a shell always raises the question of what it will run, and now the server screens commands first, so you don't have to build that step yourself. Background compaction cuts the pauses long sessions hit while summarizing, and time awareness is meant to help the model respect instructions like "timebox CI fixes to 10 minutes." Automatic system-prompt caching lowers input costs right away for products that open lots of sessions.

Claude Code is Anthropic's terminal coding agent. Version 2.1.314 adds the permissions.blockWritesOutsideWorkingDirectories setting. With it on, the Write, Edit and NotebookEdit tools refuse paths outside the working folders in every permission mode. The same version brings idle compaction to VS Code and the Code tab in Claude Desktop: a conversation over 200K tokens that sits idle gets summarized ahead of time. Version 2.1.316 adds CLAUDE_CODE_SUBAGENT_EFFORT, which sets the default effort level for subagents (helper agents that take pieces of a task), and CLAUDE_CODE_DISABLE_ULTRATHINK, which turns off the ultrathink keyword. To update:

claude update

The scariest part of leaving an agent in auto mode for hours is that it might touch your home folder or config files, and now one setting keeps its writes inside the working folder. A separate subagent effort level lets you run the main session high and the helpers doing chores low, which saves money. Org admins also get new Claude apps gateway controls, such as shared group budgets and layered policies.

Nemotron 3.5 Transcribe is an open-weight model that turns recorded speech into text. It's a mixture-of-experts (MoE) model with 34B parameters in total, of which about 4.3B are active per token. It transcribes 40 languages, translates 31 of them to or from English, and labels up to eight speakers in recordings up to an hour long. You can steer the output format or the spelling of domain terms with plain-language instructions, and it ships in BF16, FP8 and NVFP4. It loads directly in Transformers.

pip install -U transformers accelerate

NVIDIA says it posts the lowest error rates among open-weight models: a 4.71% mean WERVendor claim across the nine English sets of the Hugging Face Open ASR Leaderboard and 4.62%Vendor claim across seven languages. That means a single model can power meeting notes, call centers or medical dictation where you need speaker labels and translation together, and it runs on your own GPUs so audio never leaves your network. NVIDIA also released the Granary v2 training data and fine-tuning recipes, which makes adapting it to domain vocabulary easier.