draft Claude Code Psychosis + The Global Memory Shortage: Doug O'Laughlin, SemiAnalysis

Doug O'Laughlin says he began running SemiAnalysis’s analyst hiring case study through emerging “agentic” tools to benchmark whether they could one-shot multi-step work, and found earlier versions weren’t consistently there. He describes a noticeable shift around late December when the tool started reliably producing usable MVPs and iterating effectively, which triggered his “Claude Code psychosis” phase of trying to systematize more of his workflow.

Doug describes using Claude Code to turn his portfolio notes into structured artifacts: portfolio views, basic risk work, and an investment framework with rubrics to score positions. He emphasizes iterative loops—asking for improvements, explanations, and refinements—as the core value rather than a single perfect output.

Doug says he used Claude Code to scrape and quantify Claude Code’s footprint in public GitHub commits by searching for its sign-off signature, charting growth, and contextualizing it as a percentage of GitHub. He notes he has a cron job updating the chart daily and stresses the broader point: the marginal cost of large-scale information gathering and synthesis becomes effectively negligible (“tokens are free”).

Doug repeatedly highlights failure modes: “context rot,” increasing error rates over long sessions, and sycophancy (he calls Opus 4.6 “super sycophantic”). He says separating task execution from rubric-based evaluation can reduce bias and drift, and frames the new challenge as “hygiene”—being excellent at review and correction to safely harvest the speed gains.

Doug says “agent team” style swarms in Claude are “just an experiment” and performed poorly for him, whereas sub-agents were cleaner, and he found a Kimi “agent swarm” approach meaningfully improved performance in his testing. He frames the next frontier as chaining more verified blocks of work while keeping context windows from degrading.