Agentic Coding

This academic paper on system scaling for agentic AI directly addresses your interests in Claude Code workflows, agentic context management, and AI pair programming techniques. The paper presents a framework treating agent performance as emerging from six interacting components: reasoning substrate, memory store, context constructor, skill-routing layer, orchestration loop, and verification-and-governance layer. Rather than focusing on model scaling alone, it argues the next bottleneck is "scaling the harness"—the structured system layer around foundation models. For Claude Code users, the paper details how Claude Code implements hybrid context strategies by loading persistent project guidance (CLAUDE.md files) upfront while using just-in-time tools like grep and file reads to verify environment-dependent facts, preventing stale memory issues. On context governance, the paper identifies four critical axes: relevance, compactness, traceability, and refresh policy, emphasizing that the hard problem is not capacity but governance—constructing minimum-sufficient context through selection policies rather than fixed buffers. The framework treats prompt, skill, and memory as three temporal axes of system scaling operating at different horizons: prompts provide short-term control, skills create reusable execution patterns, and memory provides longitudinal persistence. A key contribution is proposing system-level evaluation metrics beyond one-shot success, measuring trajectory quality, memory hygiene, context efficiency, and safe evolution over time—directly applicable to evaluating your daily Claude Code and Cursor workflows. The paper also introduces CheetahClaws, a Python-native reference harness, and compares it against Claude Code and OpenClaw to make harness-level design choices explicit.

Claude Code and Cursor represent fundamentally different approaches to AI-assisted development. Claude Code operates as a terminal-native agent that autonomously handles multi-step tasks across your entire codebase, making it ideal for large refactors, automation pipelines, and headless environments where you can delegate work and review completed results. Cursor functions as an IDE-native tool built on VS Code that keeps you in control, surfacing every suggestion and change for review before application, making it better for interactive feature development, debugging, and real-time code refinement with tab autocomplete and inline editing. The choice between them depends on your workflow: use Claude Code if you work primarily in the terminal, want to hand off complex tasks to autonomous execution, or need CI/CD integration; use Cursor if you prefer staying within an IDE, want multi-model flexibility (Claude, GPT, Gemini), need full codebase indexing for context awareness, or work on teams that benefit from VS Code compatibility. Claude Code excels with very large context windows for deep codebase reasoning and strong performance on real-world coding benchmarks, while Cursor provides faster interactive feedback, immediate onboarding for VS Code users, and reduced operational risk through human-reviewed changes.