MiMo Code: Scaling coding agents to long-horizon tasks
12 hours ago
- #coding-agent
- #automated-programming
- #long-horizon-tasks
- MiMo Code is a terminal-based coding agent by Xiaomi's MiMo team, open-sourced under MIT, designed for long-horizon automated programming tasks.
- It addresses challenges in maintaining decision quality and state continuity over dozens or hundreds of steps through three themes: computation, memory, and evolution.
- For computation, Max Mode generates multiple candidate solutions per turn, using a judge to select the best, improving performance by 10–20% on SWE-Bench Pro with higher token cost.
- Goal mechanism prevents premature task termination by verifying a user-defined stopping condition, reducing infinite loop risks below 0.5%.
- Dynamic Workflow converts orchestration logic from natural language to deterministic JavaScript code for complex tasks, enabling reliable control flow and recovery from interruptions.
- For memory, a checkpoint system with independent writer subagents persists structured state files to handle context limits, allowing indefinite logical sessions via rebuild cycles.
- Memory is layered: session checkpoint, project memory, global memory, and history, with the main agent having limited write access via a notes.md scratchpad.
- For evolution, project-level memory files store cross-session knowledge, with automatic maintenance (Dream every 7 days for deduplication, Distill every 30 days for skill extraction).
- Human-in-the-loop A/B evaluations show MiMo Code's win rate exceeds 65% for tasks over 200 steps, outperforming Claude Code in complex, real development scenarios.
- MiMo Code can be installed via curl or npm, supporting multiple model access methods including a free MiMo Auto option with 1-million-token context.