Feature request for product/service
AI Models
Describe the request
The Problem
In Cursor’s current Agent and Composer workflows, a single model handles the entire lifecycle: high-level architecture, multi-file planning, brute-force file rewriting, linter execution, and syntax cleanup.
-
Frontier models (e.g., Claude Fable 5.1, GPT-6 Astra) are overkill and expensive for churning out mechanical boilerplate or streaming massive diffs.
-
Fast execution models (e.g., Composer 2.5, Grok 4.6) are cheap and fast at editing files, but prone to architectural drift, subtle logic regressions, and context blindness on complex refactors.
While users currently attempt a manual workaround (having Model A write a plan and Model B execute it), that static waterfall breaks the moment real-world execution begins.
The Proposal: Lead + Sidekick Dual-Runtime
Introduce a native dual-model execution mode (inspired by Devin’s Fusion architecture) directly selectable in the Agent model picker:
-
Lead (Architect / Supervisor): Frontier reasoning model responsible for indexing context, designing the execution path, reviewing diffs, and diagnosing unexpected runtime failures.
-
Sidekick (Executor / Worker): High-throughput, cost-efficient model dedicated strictly to generating diffs, running terminal commands, applying boilerplate, and processing routine linter passes.
Why a Dynamic Feedback Loop Beats Static “Agent-to-Agent” Planning
A real supervisor loop is fundamentally different from just splitting chats:
-
Active Supervision over One-Way Handoff: The Lead doesn’t just produce a static TODO.md and vanish; it monitors the execution cycle state-by-state.
-
Automatic Escalation on Regressions: If the Sidekick encounters breaking test outputs, missing dependencies, or API mismatches, control immediately escalates back to the Lead for structural recovery instead of having the smaller model hallucinate fixes.
-
Sanitized Context Windows: The Lead’s context remains completely clean, free from noisy compiler logs, massive raw diffs, and formatting artifacts. Only structured state transitions and architectural context reach the Lead.
-
Zero Friction: No manual prompt copying, no managing separate chat windows, and no custom sub-agent orchestration needed from the user.
Benefits for the Cursor Platform
-
Massive Token & Cost Savings: Cuts down high-tier token consumption on large codebases while maintaining frontier-grade quality.
-
Faster Throughput: Fast models stream code and diffs at double the speed without user-facing latency.
-
Higher Task Completion Rates: Combines frontier-tier system design with aggressive, high-speed execution loops.
