Cursor adds live steering for SDK agents
Long-running agents are getting more steerable, more portable, and a little more accountable. Here are three changes worth factoring into your build plans.
In today’s edition
- Cursor lets SDK agents accept mid-run steering.
- OpenAI offers opt-in API text watermarking.
- Claude cloud sessions can reach approved local folders.
Cursor lets you redirect an agent mid-run
Cursor SDK agents can now accept a new instruction while they are still working, so you can redirect a run without cancelling it and starting over.
- `run.steer()` adds a message to the agent’s next turn while the current run is active.
- If a subagent is mid-task, steering moves it to the background so it can keep working.
- Background subagents now return results as follow-up turns on the same parent run.
- Local TypeScript agents also gain custom system prompts, while MCP tool annotations can label actions such as read-only or destructive.
What it means Use steering to correct or redirect a long local Cursor run without discarding its work, while designing your own approval checks because headless tool calls can run without prompts.
Source: cursor_ai
OpenAI makes API text watermarking opt-in
OpenAI is adding text provenance controls: API customers can opt into watermarking for select models worldwide today, while eligible ChatGPT and Codex text in the EU will be watermarked over the coming weeks.
- The watermark is an invisible statistical signal that a detector can look for in generated text.
- OpenAI says watermarking did not affect model capability, speed, or how responses read in its testing.
- The signal does not identify an author, owner, account, organization, conversation, or prompt.
- Short passages may not retain a detectable watermark, and rewriting or translation can remove it entirely.
- For now, OpenAI says detector access is limited to approved researchers.
What it means Treat watermarking as an optional provenance signal for outputs that need it, not as durable attribution or a way to prove authorship once text is altered.
Source: OpenAI
Claude cloud sessions gain approved-folder access
Claude’s cloud sessions can now read and edit files in a folder you approve on your computer, while the agent and its execution environment remain in the cloud.
- Each cloud session runs in its own temporary Anthropic-managed sandbox that is destroyed when the session ends.
- The cloud session reaches local files through the Claude Desktop app, and only in folders you explicitly connect.
- The Desktop app must be online for a cloud session to access your device.
- Files opened through this bridge are processed on Anthropic’s servers rather than staying solely on the device.
- Cloud sandboxes cannot reach private, internal, link-local, or cloud-metadata addresses by default.
What it means Use cloud sessions for work that must survive a closed laptop, but treat connected folders as data sent to Anthropic and set file permissions accordingly.
Source: dfeinition