A coding agent that fine-tunes its own instincts

What it does

Momo Code is a terminal-based AI coding agent built atop SST’s opencode framework. It wraps the standard chat-and-edit workflow—supporting more than 25 LLM providers—with a self-improvement layer: a fast loop that injects learned tactics, a review gate that proposes prompt patches, and an optional training pipeline to update weights. It also ships recursive subagents, a resumable DAG engine for long-horizon tasks, and a mode where it writes Python to steer a Genesis physics simulation.

The interesting bit

The project explicitly treats agent improvement as a two-speed problem: second-level prompt injection via the KEP protocol for quick tactical gains, and hour-level weight updates via MCGS and LoRA for deeper changes. It also bundles a Genesis physics sandbox under /sim, turning the LLM into an embodied reasoning engine alongside its coding duties.

Key highlights

• Self-evolution loops: /evolve auto-selects distilled tactics via Thompson sampling; /fine-tune runs MCGS + LoRA for weight updates; /refine proposes evidence-based patches subject to human approval.
• Recursive execution: /agent spawns child processes with depth and budget rails, while /graph builds a resumable DAG of subagents that execute in parallel and retry failed nodes.
• Long-horizon autonomy: /goal injects persistent objectives into every session, and /daemon runs a foreground loop for multi-hour tasks with built-in time and spend limits.
• Simulation mode: /sim hands the agent a persistent Genesis physics world where it writes Python into a long-lived namespace, loading skills from local files.
• Claude Code interop: Inherits existing .claude/ configuration, MCP servers, and prompts, so migration is meant to be low-friction.

Caveats

• Distribution is still script-based: the published package is listed as “coming in v1.1,” so installation currently relies on a curl installer or manual build from source.
• Simulation mode needs a separate Python stack: /sim requires genesis-world installed in a Python environment, adding a dependency bridge outside the Node.js runtime.

Verdict

Worth a look if you want a local-first CLI agent that experiments with self-improvement and long-horizon planning. Skip it if you prefer a polished, all-in-one IDE experience or are wary of manually managing Node and Python environments.