Why the Best Performing Agent Harness Uses Only One Tool: Bash
sergeykarayev · x · 2026-08-22
mini-swe-agent, the top-performing framework on SWE-bench, uses only one tool: Bash. With a system prompt of just 100 lines, its minimalist design outperforms complex frameworks, sparking debate about the necessity of elaborate agent tooling.
More from coding & agent
- Vercel AI SDK Adding Support for Claude and Codex — jasonkneen · 2026-08-22
- MCP Isn't Replacing APIs: It's Changing Who APIs Are Designed For — kush_patil · 2026-08-22
- Same Model Benchmark: Agent A (45/50) vs Agent B (43/50) with Similar Cost — donk8r · 2026-08-22
- The trust crisis of AI coding: "Claude said it was fine" — uwukko · 2026-08-22
- GitHub Repo Lists LLM Skills for Claude, Gemini, and Custom Agents — tom_doerr · 2026-08-22
- A SKILL.md that forces Cursor and Claude into Fred George's programming style — cti97 · 2026-08-22