TL;DR

Cursor 3 launched April 2, 2026 with agent-first architecture. Agents Window replaces Composer pane for parallel agent execution. Composer 2.0: 61.3 CursorBench (39% improvement), 200+ tok/s via custom GPU kernels. Background and Cloud Agents enable autonomous coding.

Key Facts

• Who: Anysphere
• What: Cursor 3 with agent-first architecture, Composer 2.0, Cloud Agents
• When: April 2, 2026
• Impact: IDE evolution from autocomplete to autonomous orchestration

What Changed

Anysphere released Cursor 3 on April 2, 2026, replacing the Composer pane with a standalone Agents Window for orchestrating multiple AI agents in parallel.

Composer 2.0 specifications:

Metric	Value	Delta
CursorBench	61.3	+39% vs v1.5
Throughput	200+ tok/s	~3x vs Opus 4.7 API
TTFT	150ms	~3x vs Opus 4.7 API

Technical optimizations: custom MLA kernel (1.7x faster attention), speculative decoding (65% hit rate), prompt prefix caching.

Two execution modes:

1. Background Agents: Tasks continue when laptop closed; cloud handoff for 30+ minute tasks
2. Cloud Agents: Autonomous workflows on Anysphere cloud ($200/month Max tier)

According to Cursor’s blog, agents are now primary workflow orchestrator, not auxiliary panel.

Why It Matters

The agent-first shift changes developer interaction with AI coding tools:

• Parallel execution: Multiple agents in isolated git worktrees — refactoring, testing, documentation simultaneously
• Autonomous workflows: Background/Cloud Agents eliminate synchronous interaction bottleneck
• IDE-native latency: 200+ tok/s provides tangible UX advantage over API alternatives

Market: Cursor 3 (18% share) vs Claude Code (18%, $20/month) vs Windsurf (8%). Cursor’s VS Code fork offers extension maturity that terminal-first Claude Code lacks.

🔺 Scout Intel: What Others Missed

Confidence: high | Novelty Score: 82/100

Coverage focuses on benchmarks. Deeper signal: Anysphere positions Cursor as orchestration layer for autonomous coding agents — first commercially viable “code while you sleep”. TokenMix data shows CursorBench rewards IDE-native behaviors; Opus 4.7 still leads SWE-Bench Verified (87.6% vs ~80%). Composer 2 best inside Cursor; Opus 4.7 superior for headless API automation.

Key Implication: Teams choosing Cursor 3 vs Claude Code should evaluate workflow preference (VS Code vs terminal), not model quality. Both support autonomous agents; differentiation is IDE form factor.

What This Means

VS Code-native developers: Cursor 3 offers highest capability in IDE form factor. Agent-first architecture transforms AI to workflow orchestrator. Parallel agent UI and Ghost Mode unique.

Terminal-centric workflows: Claude Code better fit. Cursor cannot replicate terminal-native workflow. Claude Code Routines provide equivalent background execution.

What to watch: Composer 3 unannounced (Q4 2026 likely). Composer 2 provides material jump now. SpaceX-Cursor acquisition ($60B rumor) could shift default to Grok variant.

Sources

• Cursor Blog: Cursor 3 Release — Anysphere, April 2026
• InfoQ: Cursor 3 Agent-First Interface — InfoQ, April 2026
• TokenMix: Cursor Composer 2 Review — TokenMix, April 2026