Cursor SDK Lets Developers Redirect AI Agents Mid-Task

Cursor has launched version 1.0.31 of its npm package @cursor/sdk and PyPI package cursor-sdk, introducing a way to redirect AI agents during active runs. The update features a new run.steer(text) method that injects instructions directly into an in-flight agent turn. This allows developers to correct an agent's path without canceling its active work or discarding its accumulated state. The method resolves with either complete_delivered, meaning the instruction was successfully injected, or revert_to_followup, which signals that the application should queue the instruction as a normal follow-up turn. If a foreground subagent is busy when steered, the system automatically moves it to the background so it can continue working while the parent agent processes the new direction.

This release also addresses the issue of disappearing background subagent results. Previously, subagent outputs could be lost if the parent turn ended first. Now, version 1.0.31 returns completed subagent results to the parent as follow-up turns on the same run. Developers can use run.stream() to yield events through these follow-up turns, while run.wait() will resolve only after all subagent tasks finish. This means a single wait operation can now cover both the parent turn and any parallel subtasks, though developers must ensure their integrations keep consuming the stream until the run officially closes.

Additionally, the update brings Model Context Protocol metadata to custom tools via an annotations field. Developers can now pass hints like readOnlyHint, destructiveHint, idempotentHint, and openWorldHint to help the model reason about tool side effects before selecting them. For security, developers should still enforce authorization inside tool handlers and use the disallowedTools option introduced in version 1.0.27. Finally, a new systemPrompt option on Agent.create() and Agent.resume() lets developers replace Cursor's built-in system prompt for local TypeScript agents while preserving existing rules, skills, and tool schemas.