# GenAlphAI

> GenAlphAI is a research-driven AI publication for engineers and operators: deep, evidence-backed analysis of agentic systems, model evaluation, and the economics of AI software.

## Pillars

- [Agent Harness Engineering and Agentic Loops: 2026 Field Guide](https://genalphai.com/agentic-loops-and-harness-engineering/): The agent harness, not model IQ, decides whether autonomous coding agents ship. How ReAct, Plan-Execute-Verify, and the Ralph Wiggum loop really work.
- [Generative Engine Optimization: How to Earn AI Citations](https://genalphai.com/generative-engine-optimization-guide/): Generative engine optimization decides whether ChatGPT and AI Overviews cite you. The 2026 playbook: crawlers, llms.txt, and AI share of voice.
- [AI Coding Agent Economics: Real ROI and Cost per Pull Request](https://genalphai.com/economics-of-ai-coding-agents/): AI coding ROI, demystified: real cost per pull request ($1-$30 in tokens, $20-$80 all-in), why the 4:1 claim doesn't hold up, and when local-first agents beat the cloud.
- [SWE-bench Pro vs Verified: Can You Trust Coding Benchmarks?](https://genalphai.com/swe-bench-pro-vs-verified/): SWE-bench Verified was deprecated after 59.4% of its hard tasks had flawed tests. What SWE-bench Pro and the DeepSWE audit reveal about coding agent benchmarks.

## Articles

- [Context Rot and the Dumb Zone: Engineering Past 100k Tokens](https://genalphai.com/context-rot-and-the-dumb-zone/): Context rot degrades agent accuracy well inside the advertised context window. Here's the evidence for the 100k dumb zone and the architecture that fixes it.
- [AGENTS.md vs CLAUDE.md vs Cursor Rules: Config Done Right](https://genalphai.com/agents-md-vs-claude-md/): AGENTS.md, CLAUDE.md, and .cursor/rules compared: ownership, the three-tier permission model, context window budgeting, and the canonical-plus-adapters pattern.
- [The Ralph Wiggum Loop: Why Stateless Agents Beat Smart Ones](https://genalphai.com/ralph-wiggum-loop-stateless-agents/): The Ralph Wiggum loop, coined by Geoffrey Huntley in July 2025, restarts a stateless agent every iteration with git as memory. Here's why it works.
- [Reasoning-First LLMs: Make Models Reason, Not Rationalize](https://genalphai.com/reasoning-first-llms/): Reasoning-first LLMs require process supervision, self-consistency, and faithfulness probes, because chain of thought often rationalizes answers post hoc.

## Important Links

- [llms-full.txt](https://genalphai.com/llms-full.txt): Consolidated Markdown archive of all articles.
- [RSS](https://genalphai.com/rss.xml)