We're All Addicted To Claude Code

Calvin French-Owen (Segment co-founder, ex-OpenAI Codex) breaks down why Claude Code beats IDEs and how coding agents will reshape software teams.

• Claude Code’s key edge: it spawns sub-agents using Haiku to explore the filesystem in parallel, splitting context more effectively than IDE-based tools.
• Cursor uses semantic/embedding search for context; Claude Code and Codex just use grep — and grep wins because code is dense and low-noise.
• Codex runs compaction after every turn so it can sustain 24–48 hour jobs; Claude Code does not compact mid-session, making it better for shorter interactive work.
• Context poisoning is real: Calvin clears context when it exceeds ~50% of the token window to avoid degraded output quality.
• GEO (generative engine optimization) already works: a competitor published a biased top-5 list with themselves ranked first, and LLMs now recommend it uncritically.
• Supabase dominates LLM recommendations for backend setup because its open-source docs saturated the training corpus — beating paid tools with better documentation alone.
• Segment’s original integration-wiring value has dropped to near zero with coding agents; the durable value is now the data pipeline, audience management, and customer orchestration layer.
• Senior engineers benefit most from coding agents because the bottleneck shifts to architectural taste and knowing which changes to kick off, not typing speed.

2026-02-06 · Watch on YouTube