Harnesses beat raw agents: HyperFrames cuts video-agent token spend 30% vs Claude Code
Efficient-Clock337 · reddit · 2026-11-16
An engineer who can't edit videos built harnesses around coding agents (Claude Code, Codex, Grok) and now runs two fully agent-edited short-video pipelines. Citing HyperFrames' engineering writeup on a DeepSeek Harness-based video agent: without changing the frontier model, better scaffolding cut token spend 30% and turns 35% vs Claude Code. Key ideas: models provide intelligence, harnesses provide domain competence; package domain skills as swappable benchmarkable plugins; review session traces to convert stall points into deterministic tools. Automating a vertical is about building the right harness, not training custom models.
Related event: Coding Agents With Custom Harnesses Cut Video Editing Token Use 30%(2 posts)→
More from coding & agent
- YC F26's FlashDrive launches a shared cloud drive for AI agents — damnGruz · 2026-11-16
- Reddit dev details 7 chronic failures of AI agents, from endless planning to fake progress — Nmoreland51 · 2026-11-16
- Open-source coding.sh measures where your coding agents waste time and money — AsianKidTV · 2026-11-16
- 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