I built a coding agent in ~970 lines of Python and benchmarked it honestly

Short summary

The author built nano-harness, a ~970-line Python coding agent with one loop and three tools (bash, read_file, edit_file), scoring 59.6% on Terminal-Bench 2.0 with Claude Opus 4.8. Key design principles include an evidence-gated 'done' verification system that challenges the model to prove task completion before accepting results. The project demonstrates that a minimal, readable harness can put up real benchmark numbers without embeddings, planners, or sub-agents.

• •Built a 970-line coding agent harness scoring 59.6% on Terminal-Bench 2.0
• •Core design: one loop, three tools, evidence-gated completion verification
• •Benchmark progression from 20% (Haiku) to 59.6% (Opus) with hardening

Generated with AI, which can make mistakes.