LoongFlow loongflow

PEES (Plan-Execute-Evaluate-Summary) iterative problem-solving methodology with LoongFlow engine for complex tasks. Use when tasks need structured iteration, optimization, evolution, or when user mentions loongflow/PEES/PES.

install
source · Clone the upstream repo
git clone https://github.com/baidu-baige/LoongFlow
Claude Code · Install into ~/.claude/skills/
T=$(mktemp -d) && git clone --depth=1 https://github.com/baidu-baige/LoongFlow "$T" && mkdir -p ~/.claude/skills && cp -r "$T/.claude/skills/loongflow" ~/.claude/skills/baidu-baige-loongflow-loongflow && rm -rf "$T"
manifest: .claude/skills/loongflow/SKILL.md
source content

LoongFlow — PEES Iterative Problem Solving

Use this skill when the user wants to iteratively improve a solution — optimization, evolution, structured retries with learning, or any task that benefits from multiple rounds of refinement rather than a one-shot attempt.

Step 1: Analyze and Advise

Before starting, analyze the task and advise the user on which mode to use. Present both options clearly:

Native PEES (recommended for simple tasks):

  • Best for: single-file fixes, small features, bug fixes, focused improvements
  • How it works: You run Plan-Execute-Evaluate-Summary iterations yourself within this conversation
  • Pros: Fast, no setup, no external dependencies, transparent workspace with full history
  • Cons: Limited to ~5 iterations, single-threaded, no population-based evolution

LoongFlow Engine (recommended for complex tasks):

  • Best for: optimization problems, multi-file projects, tasks needing many iterations (50+), population-based evolution with diversity preservation
  • How it works: Downloads the LoongFlow framework, creates a
    general_agent
    task, runs evolutionary optimization in the background, monitors via cron
  • Pros: Powerful evolutionary engine with multi-island model, Boltzmann selection, MAP-Elites diversity, checkpointing, cost tracking
  • Cons: Requires
    ANTHROPIC_API_KEY
    and
    ANTHROPIC_BASE_URL
    , setup time, runs as background process
  • Source: https://github.com/baidu-baige/LoongFlow

Ask the user which mode they prefer before proceeding.

Step 2: Follow the Mode Guide

Once the user chooses, read the corresponding reference file for detailed instructions:

  • Native PEES → Read
    references/native-pees.md
    and follow it
  • LoongFlow Engine → Read
    references/engine-mode.md
    and follow it

Architecture Reference

LoongFlow supports three tiers for agent projects:

TierDescriptionBest For
SimpleReAct loop + persistent memoryChatbots, tool calling, format conversion
StandardReAct + self-evaluation + iterative improvementCode review, document generation, data analysis
AdvancedPEES evolution loop with loongflow-memoryMath optimization, algorithm design, NP-hard problems

Complexity Assessment

Task Analysis
├── Only needs conversation + simple tools? → SIMPLE
├── Needs file operations or code generation?
│   ├── Has numerical evaluation metric? → ADVANCED
│   └── No numerical metric? → STANDARD
└── Needs iterative optimization?
    ├── Has clear scoring function? → ADVANCED
    └── Qualitative improvement? → STANDARD