Cursor SDK Adds Live Agent Steering and Background Subagent Results
Cursor SDK 1.0.31 returns background subagent results to the parent run and adds MCP tool annotations for custom tools. Live steering with run.steer() and replaceable system prompts are documented in the TypeScript SDK reference for local agents. As of October 7, 2026, the changelog is at 1.0.35 and 1.0.32 changed how steering behaves while an agent waits on background subagents.
Cursor SDK 1.0.31 changes what happens to work an agent delegates: background subagent results now return to the parent run instead of being dropped at the end of the parent turn, and custom tools can carry MCP annotations. Alongside it, Cursor's TypeScript SDK reference documents two controls for agents that are already running or about to start: run.steer() for a local TypeScript run, and a replaceable system prompt on Agent.create(). As of October 7, 2026, the SDK changelog lists only the subagent and annotation items under 1.0.31; the steering and system prompt sections live in the TypeScript SDK reference, which carries the runtime limits for each (Cursor on X, follow-up, follow-up, follow-up, follow-up).
The useful headline for developers building with the Cursor SDK is that runs can now accept input while work is in progress, while delegated work can finish in the background and still report back through the original run. The new behavior is not uniform across runtimes, so check the language and local-versus-cloud limits before changing an integration.
Steering a Running Agent
The TypeScript SDK has run.steer(text). Cursor's announcement says the method steers an agent while it runs; the TypeScript SDK reference specifies that the message is injected into the turn that is already running instead of waiting for it to finish, and that the promise settles once the turn confirms whether it appended the text: complete_delivered means the turn has the message, revert_to_followup means it did not take it, so deliver it with agent.send() after the run. If a foreground subagent is running when you steer, it moves to the background and keeps running so the parent turn can take the message. Steering is for local runs only: cloud runs and detached local handles expose steer but always resolve revert_to_followup. steer is optional on Run, so check run.steer before calling it (Cursor TypeScript SDK reference; Cursor's announcement).
SDK 1.0.32 changed one case. Before it, run.steer(text) resolved revert_to_followup when the agent had ended its turn to wait on background subagents, so the message waited until that work finished; now the steer runs right away as the next turn, and the background results still arrive afterward. TypeScript local runs only (Cursor SDK changelog, 1.0.32 entry).
That gives a local integration a way to redirect active work without cancelling the whole run. The response value matters: code should handle revert_to_followup rather than assume that every steering message was delivered into the active turn.
Background Subagents Return Their Results
Before this release, a background subagent's result was dropped when the parent turn ended. With 1.0.31, the subagent's result returns to the parent as a follow-up turn on the same run. run.stream() keeps yielding through those turns, and run.wait() resolves after them. Cursor lists this behavior for local agents in both TypeScript and Python (Cursor SDK changelog; Cursor on X).
For developers, this makes background delegation easier to consume through the run's existing stream or wait path: the child result becomes part of the parent run's completion rather than a separate result channel.
Custom Prompts and MCP Tool Annotations
Two additional changes give developers more control over a local agent's setup:
- Replace the system prompt. TypeScript agents can set
systemPromptonAgent.create()to replace Cursor's built-in system prompt for the main agent loop with your own text. The model loses the coding-assistant identity, tool-use protocol, and communication guidance, so restate whatever the agent still needs; tool schemas, rules, and skills still load, and subagents keep their own prompts. The override is for local agents only (combining it withcloudthrows aConfigurationError), must be non-empty, is not persisted on the agent, so pass it again onAgent.resume(), and access is enabled per account: without it, the firstsend()fails with an error naming--system-prompt(Cursor TypeScript SDK reference; Cursor on X). - Pass MCP annotations to custom tools. A TypeScript
local.customToolsentry can includeannotations, includingtitle,readOnlyHint,destructiveHint,idempotentHint, andopenWorldHint. Cursor says these are descriptive hints passed to the model; the SDK does not enforce them (Cursor SDK changelog; Cursor on X).
The practical caveat: steering and prompt replacement are limited to local TypeScript agents, while background results are supported for local TypeScript and Python agents. Tool annotations are not a security boundary. Treat them as model-facing descriptions, and keep actual authorization and safety checks in the host application.
What to Check Before Updating
| Capability | Runtime and language in Cursor's docs | Important limit |
|---|---|---|
run.steer() | Local TypeScript (SDK reference) | Cloud runs and detached local handles resolve revert_to_followup; since 1.0.32 a steer sent while the agent waits on background subagents runs as the next turn |
| Background subagent results | Local TypeScript and Python | Results return as follow-up turns on the same run |
Custom systemPrompt | Local TypeScript (SDK reference) | Pass it again on Agent.resume(); access is enabled per account; cloud plus systemPrompt throws ConfigurationError |
| MCP annotations on custom tools | TypeScript | Descriptive hints only; not enforced by the SDK |
If your integration is already on the Cursor SDK, the most actionable first step is to check the changelog, then add explicit handling for steering fallback and for the extra follow-up turns produced by background subagents.
Corrections and Updates
- October 7, 2026: this page said Cursor's SDK changelog documented
run.steer()and thesystemPromptoverride under 1.0.31. As of October 7, the changelog's 1.0.31 entry lists only background subagent results and custom tool annotations; the steering and system prompt sections, with the same local-only limits, are in the TypeScript SDK reference. The 1.0.32 change to steering while waiting on background subagents was added.
Sources
- Cursor, SDK changelog (release notes): https://cursor.com/docs/sdk/changelog
- Cursor, TypeScript SDK reference (Steering a run in flight; Replacing the system prompt): https://cursor.com/docs/sdk/typescript
- Cursor on X, agent steering announcement: https://x.com/cursor_ai/status/2107141004482793827
- Cursor on X, background subagent results: https://x.com/cursor_ai/status/2107141021947904430
- Cursor on X, MCP annotations: https://x.com/cursor_ai/status/2107141038427308473
- Cursor on X, custom system prompts: https://x.com/cursor_ai/status/2107141054071968154
- Cursor on X, TypeScript and Python support details: https://x.com/cursor_ai/status/2107141066403275081
Read next
Related coverage- Cursor
Cursor SDK: Programmatic Cursor Agents in TypeScript and Python, Local or Cloud
Cursor launched its SDK in public beta on April 29, 2026 as a TypeScript package. As of October 9, 2026 it ships for TypeScript (`@cursor/sdk`) and Python (`cursor-sdk`), released together at version 1.0.35, and the docs no longer call it beta. Agents run locally, on Cursor-hosted VMs, or on self-hosted workers.
- Cursor
Cursor SDK Bridge Opens Agent Control to Rust, Go, and More
Cursor has open-sourced an SDK Bridge that exposes a stable sdk.v1 protocol for driving Cursor agents from Rust, Go, Java, and other languages. Adapters run a small local bridge and communicate over Connect RPCs, while the bridge handles the connection to Cursor's API.
- Cursor
Cursor Agent SDK: Build, Evaluate, and Deploy Agents as Code
Cursor's Agent SDK gives teams a code-first way to define, test, evaluate, and deploy AI agents. Its docs describe project instructions, tools, and evaluations as versionable files, with Markdown and TypeScript in the project. Developers can use Cursor-managed hosting or self-hosted infrastructure.
- Cursor
You Can Now Delegate Tasks to Cursor From Inside Notion
Notion used the Cursor SDK to embed coding agents directly in its workspace. You can now tag Cursor in a doc, mention it in a thread, or assign it a database issue, and Cursor plans, builds, tests, verifies its own work, and opens a PR. Notion says it built the integration in a few weeks on the Cursor SDK.
Frequently Asked Questions
What did Cursor add to the SDK?
As of October 7, 2026, the changelog's 1.0.31 entry lists background subagent results that return to the parent run (local TypeScript and Python agents) and MCP annotations for custom tools (TypeScript). Live steering with run.steer() and replaceable system prompts are documented in the TypeScript SDK reference for local agents.
Can I steer a Cursor SDK cloud run?
No. Cursor's TypeScript SDK reference says run.steer() works on local runs only; cloud runs and detached local handles expose steer but always resolve revert_to_followup. Deliver the message with agent.send() after the run instead.
Do MCP annotations enforce permissions?
No. Cursor describes them as descriptive hints passed to the model. The SDK does not enforce them.
Get an email when Cursor changes its plans, prices or limits
We check every change against Cursor's own pages before we send. You also get the weekly AI Catchup; one click unsubscribes.