Models give intelligence, harnesses give competence: coding agents cut video-editing tokens 30%
Efficient-Clock337 · reddit · 2026-11-16
An engineer who can't edit videos shares how he built harnesses around coding agents (Claude Code, Codex, Grok) to run two fully automated short-video pipelines. Citing HyperFrames' engineering writeup: their video agent on a DeepSeek harness beat general-purpose agents like Claude Code with 30% less token spend and 35% fewer turns — by changing scaffolding, not the model.
Key takeaways:
- Models provide intelligence, harnesses provide domain competence: general agents waste tokens writing throwaway tooling and guessing visual/audio alignment; deterministic tools that measure instead of guess change everything
- Keep capabilities modular: package domain skills as standalone plugins to benchmark, swap, and evolve each time a new frontier model drops
- Trace-driven improvement: review session traces step-by-step, find stall/loop points, convert them into deterministic plugins
Conclusion: automating a new vertical isn't about training custom models — it's about building the right harness with plugins, context, and feedback loops.
Related event: Coding Agents With Custom Harnesses Cut Video Editing Token Use 30%(2 posts)→
More from coding & agent
- Dev ships MCP server for a trading journal so Claude can query PnL and win rate — Occ_inv · 2026-11-16
- 10,000-Word Playbook: Automating Overseas KOL Marketing with AIsa and AI Agents — sujingshen · 2026-11-16
- Economist builds TeXugo, a free open-source LaTeX editor for academic papers, with Claude Code — joshgans · 2026-11-16
- Full YouTube walkthrough of solo operator's Business Brain AI workspace — aryanXmahajan · 2026-11-16
- Autopilot Browser ships MCP Server for natural-language browser automation — modelcontextprotocol · 2026-11-16
- KDP Niche Research MCP sells Kindle demand/BSR/revenue data via x402 USDC on Base — modelcontextprotocol · 2026-11-16