Cursor 3 Launches With a Unified Agent-First Workspace

Cursor transitioned its coding environment into a unified agent workspace today with the release of Cursor 3. The platform recently reached $2.0 billion in annualized revenue with over 1 million daily active users. Following a $2.3 billion Series D that pushed the company’s valuation to $29.3 billion, this release targets enterprise development workflows where context retention across autonomous tasks is critical.

Unified Agent Workspace

The core of Cursor 3 is a combined interface that merges distinct interaction models into a single session. Users now switch between Ask, Edit, and Agent modes without losing contextual history. Ask replaces the traditional Chat function for debugging and codebase queries. Edit supersedes Composer, handling multi-file generation under direct developer supervision.

The new Agent mode executes fully autonomous tasks. It runs shell commands, processes large-scale refactoring operations, and initiates web searches without manual prompting. If you build multi-agent systems, you understand the difficulty of maintaining state across different reasoning modes. Cursor handles this by keeping the context window unified across all three interaction types.

Composer 2 Benchmarks and Architecture

Powering these features is Composer 2, an in-house model designed specifically for coding tasks. Cursor states this model matches GPT-5.4 performance and outperforms Anthropic’s Opus 4.6 on coding benchmarks. The operational advantage lies in inference efficiency. Composer 2 runs at one-tenth the cost of GPT-5.4.

Model	Target Use Case	Relative Cost
Composer 2	Agentic Code Generation	0.1x
GPT-5.4	General Multi-step Reasoning	1.0x

The model’s origin has prompted scrutiny. Researchers at Moonshot AI state Composer 2 uses their Kimi K2.5 model without license compliance. They identified identical tokenizers and model IDs (kimi-k2p5-rl-0317-s515-fast) operating within the Cursor environment. Cursor has not addressed these licensing claims.

Context Integrations and Security

Cursor 3 introduces deep support for the Model Context Protocol. Developers configure third-party resources via a local .cursor/mcp.json file. A new YOLO mode permits agents to execute tool calls automatically without requiring manual developer approval for each step.

The editor now automatically appends terminal outputs and Git commits to the active context window. For organizations protecting proprietary code, Cursor added Self-Hosted Cloud Agents. This deployment model runs agents entirely within the customer’s network infrastructure. It keeps codebase data, build outputs, and secrets isolated. A supporting system of .cursorignore and .cursorindexignore files prevents specific directories from ever reaching an LLM or local index.

The system also supports always-on agents. These spin up in isolated cloud sandboxes based on external event triggers. Developers can route alerts from tools like PagerDuty, Slack, Linear, or GitHub directly to an agent for automated triage and pull request creation. If you compare this against other AI coding assistants, the emphasis has shifted from localized code generation to broader workflow automation.

Migrating to Cursor 3 requires updating your project configurations to control agent autonomy. You should define your .cursorignore boundaries immediately to prevent unintended file indexing before testing the autonomous Agent mode. Evaluate the YOLO mode for MCP tools in a sandboxed directory first to understand how aggressively Composer 2 executes shell commands without supervision.