Fancy Agent Harnesses Are Dead Weight

🔥 What happened A new arXiv paper (Brilliantov et al., Sep 30, 2026) finds that under equal time budgets and the same frontier LLM, elaborate open-source agent harnesses deliver zero advantage over a single minimal coding agent with read/write/bash primitives. 💡 Why it matters The backbone model drives performance, not the orchestration. Large-scale ablations show multi-agent orchestrators and retrieval subagents are redundant in coding settings — teams burning months on harness engineering get no benchmark lift for the effort. ⚡ Our take The harness industry just lost its expensive hobby. Anyone pitching subagent architectures in 2026 should spend that budget on the next frontier model instead.