Xiaomi's MiMo Code claims it beats Claude Code past 200 steps

Abstraction: Long-horizon coding agent endurance gap and competing harness approaches

Key points:

• Xiaomi's open-source MiMo Code claims to outperform Claude Code on tasks exceeding 200 steps; benchmark is self-reported from internal beta and 576-developer survey
• Three common long-run failure modes: hypothesis lock-in (wrong assumption patched repeatedly), error compounding, and context drift
• Berkeley's Agents' Last Exam (Dawn Song, UC Berkeley RDI lab): grades finished artifacts, not demos; GPT-5.5 + Codex scored below 50% on easiest tier and under 10% on hardest; Claude Code near-zero on hardest tier
• Three harness approaches compared: Claude Code nested subagents (frontier planner + cheaper executors, capped 5 levels deep); Arbor from Renmin University (persistent hypothesis tree with checkpoint/resume, 86.36% Any-Medal on MLE-Bench Lite); Xiaomi MiMo Code (terminal-native, 200+ step focus)
• "Endurance gap" = steps an agent survives before dropping the task; now becoming a procurement evaluation criterion alongside coding leaderboard scores

Connections: Xiaomi · Claude Code · Anthropic · Agentic Coding · AI Agents

Source: https://thenewstack.io/coding-agent-endurance-gap/