Long-running Agents in Research Preview

Source: Cursor Date: 2026-02-12 URL: https://cursor.com/changelog/02-12-26

Summary

Cursor ships long-running agents in research preview (plan-first execution, autonomous operation over extended task horizons, improved PR completeness). Available via cursor.com/agents on Ultra, Teams, and Enterprise plans. Internal testing showed these agents complete work “previously too hard for regular agents.”

Implications

Agent-IDE feature race. Long-running agents with plan-first execution are Cursor’s most direct response to Devin and SWE-agent benchmarks. The framing — “fewer follow-up items” and “previously too hard” — signals these are being positioned against the frontier of autonomous coding capability, not incremental improvements. Gating to Ultra/Teams/Enterprise only implies high compute cost, consistent with the isolated VM model used for cloud agents.

Pricing model evolution. Gating long-running agents to paid tiers while background agents are universal creates a clear capability tier: standard agent (everyone), background agent (everyone with cloud), long-running agent (paid). This is the pricing ladder that justifies premium plan pricing as the autonomous capability ceiling rises.

Watch: Whether long-running agents exit research preview with transparent task duration/cost metrics and whether the “plan first” approach generates auditable artifacts (plans saved as files, visible in the dashboard).