Raw transcript — Claude Code 2.0 Agentic Coding: No, other agents aren’t even close.

Source: https://www.youtube.com/watch?v=nGhsgdQplHw Duration: 28m Captured: 2026-04-20

Full clean transcript stored at /tmp/backfill-2026-04-20/claude-code-2-0.txt during ingestion (5,632 words). Per copyright policy, raw transcript preserved for internal reference. Re-download via:

yt-dlp --write-auto-sub --sub-lang en --skip-download --sub-format vtt -o "/tmp/yt-process/%(id)s" "https://www.youtube.com/watch?v=nGhsgdQplHw"
python3 ~/.claude/scripts/vtt-to-text.py /tmp/yt-process/nGhsgdQplHw.en.vtt

Key segments (timecoded)

• [00:00–02:00] Two parallel demos — in-loop on local device (migrate codebase to Claude Agent SDK) AND out-of-loop on dedicated agent device (build OpenAI realtime prototypes). /AFK agents kicked off with reusable MCP server prompt.
• [02:00–05:00] New Claude Code feature: chained custom slash commands. /scout-plan-build composes three sub-prompts. Agentic prompts as composable units.
• [05:00–07:00] Scout step delegates file searching from the planner. Multi-vendor sub-agents (Gemini Lite, Codex, Gemini Flash, Haiku) for diversified perspectives. R&D framework: “Reduce and Delegate” — the only two ways to manage context window.
• [07:00–10:00] Planner reads scout’s relevant-files output. Core Four reiterated: context, model, prompt, tools. Eight applications migrated back-to-back-to-back without context blowup.
• [10:00–13:00] Critical context warning — single primary agent ate 51% of its window. ADWs are the structural fix.
• [13:00–15:00] Auto-compact buffer issue — Claude Code 2.0’s auto-compact eats 22% of context window by default. Recommendation: turn it off. Restores full 91% free.
• [15:00–18:00] Custom output style demoed (observable tools + diffs + TTS). Concern that Claude Code is being made “the tool for everyone, not just engineers.”
• [18:00–22:00] Out-of-loop dedicated agent device review. M4 Mac Mini ran the OpenAI agent SDK prototypes end-to-end while the engineer was in-loop on something else. Sleep-60s-check-status pattern.
• [22:00–27:55] Sponsorship — Tactical Agentic Coding course pitch. “Build the system that builds the system.” 8-lesson core + Agentic Horizon extended lessons. Pricing tiers, early bird, principal-AI-coding-member discount. Closing: “Claude 4.5 Sonnet is the best coding model in the world.”