Same model, better harness: Opus 5.5 jumps from 8/20 to 16/20 accepted tasks
Arindam_1729 · x · 2026-11-17
Developer Arindam shares that with the same model, same prompts, and same 20 real tasks, one weekend of harness engineering took Opus 5.5 from 8/20 to 16/20 accepted tasks.
Key numbers:
- Retry loops: 5 → 0
- Cost per accepted task: $4.10 → $1.60
- Daily babysitting: 2 hours → 25 minutes
What changed: a done-when condition, a 40-line map file, a test gate, a 3-retry cap, and rules defining what the agent can touch. The author notes a longer prompt fixes just one layer — the other four must be built around the model. Details in his companion guide, plus a longer article on proactive agents.
More from coding & agent
- 7 Claude Computer Use prompts to fix your slow PC like a $100/hr IT tech — TawohAwa · 2026-11-17
- Anthropic: one general agent often beats a sprawl of specialized subagents — timigod · 2026-11-17
- His AI agent secretly co-founded a startup with another agent and became its CTO — alex_verem · 2026-11-17
- unslop: a Claude skill that de-slops vibecoded UI hits 17.5k GitHub stars — socialwithaayan · 2026-11-17
- KORMo-Web-11B, a Korean VLM for GUI and web agents, trends on Hugging Face — KORMo-VL · 2026-11-17
- User builds fully automated video creation pipeline with Grok Bot, friend gets hooked — XFreeze · 2026-11-17