120-run experiment compares session strategies for coding agents

Sunny Code's engineering blog tested five coding-agent session strategies (fresh, continue, compact, fork, fresh+summary) across two models in 120 runs, 102 successful. No universal winner emerged: full history raised Sol's cost to $2.29 per run versus $0.19 for fresh, while DeepSeek's continue runs were faster.

• 120 runs: 5 strategies × 2 models × 2 rounds × 3 repetitions
• Sol: continue $2.29 and 14 min vs $0.19 and 2.7 min for fresh
• DeepSeek: fresh 6.8 min, continue 3.0, fork 2.2 — all 3/3
• Fresh+summary passed functional checks 4/6 vs 6/6 for fresh