Latest deep dive
Generative Engine Optimization: How to Earn AI Citations
Search is becoming synthesis. If ChatGPT, Perplexity, and Google's AI Overviews don't cite you, you're invisible, and the playbook is not the SEO playbook you already know.
Pillars
Pillar
Generative Engine Optimization: How to Earn AI Citations
Generative engine optimization decides whether ChatGPT and AI Overviews cite you. The 2026 playbook: crawlers, llms.txt, and AI share of voice.
June 11, 202616 min
Pillar
AI Coding Agent Economics: Real ROI and Cost per Pull Request
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.
June 11, 202620 min
Pillar
SWE-bench Pro vs Verified: Can You Trust Coding Benchmarks?
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.
June 10, 202618 min
More analysis
AI Coding Agent Economics: Real ROI and Cost per Pull Request
Frontier labs now ship more AI-written code than human-written code, but the viral ROI numbers are wrong. Here is the money math that survives CFO scrutiny. 20 min Context Rot and the Dumb Zone: Engineering Past 100k Tokens
Bigger context windows didn't fix attention. Past roughly 100k tokens agents get lost in the middle, and the fix is architectural, not bigger. 11 min SWE-bench Pro vs Verified: Can You Trust Coding Benchmarks?
OpenAI deprecated the benchmark everyone quoted, an audit found graders wrong on a third of verdicts, and frontier models got caught reading the answer key. Here is what actually measures a coding agent in 2026. 18 min AGENTS.md vs CLAUDE.md vs Cursor Rules: Config Done Right
The config files are your agent's control plane. Get the three-tier permission model and context budgeting right, or watch instruction adherence rot. 9 min The Ralph Wiggum Loop: Why Stateless Agents Beat Smart Ones
Wiping the agent's memory every iteration sounds like sabotage. It's actually the most reliable way anyone has found to run a coding agent for hundreds of turns. 9 min Reasoning-First LLMs: Make Models Reason, Not Rationalize
Your model's chain of thought is a narrative, not a derivation. Here is the stack that forces it to actually compute the answer. 11 min
Frontier labs now ship more AI-written code than human-written code, but the viral ROI numbers are wrong. Here is the money math that survives CFO scrutiny. 20 min Context Rot and the Dumb Zone: Engineering Past 100k Tokens
Bigger context windows didn't fix attention. Past roughly 100k tokens agents get lost in the middle, and the fix is architectural, not bigger. 11 min SWE-bench Pro vs Verified: Can You Trust Coding Benchmarks?
OpenAI deprecated the benchmark everyone quoted, an audit found graders wrong on a third of verdicts, and frontier models got caught reading the answer key. Here is what actually measures a coding agent in 2026. 18 min AGENTS.md vs CLAUDE.md vs Cursor Rules: Config Done Right
The config files are your agent's control plane. Get the three-tier permission model and context budgeting right, or watch instruction adherence rot. 9 min The Ralph Wiggum Loop: Why Stateless Agents Beat Smart Ones
Wiping the agent's memory every iteration sounds like sabotage. It's actually the most reliable way anyone has found to run a coding agent for hundreds of turns. 9 min Reasoning-First LLMs: Make Models Reason, Not Rationalize
Your model's chain of thought is a narrative, not a derivation. Here is the stack that forces it to actually compute the answer. 11 min