FlagScale-Agent

---

🌟 Overview

FlagScale-Agent is an autonomous AI agent specialized in large-scale distributed training, inference, and serving infrastructure. Built on the ReAct (Reasoning + Acting) paradigm, it combines LLM reasoning with domain-specific tools, knowledge, and safety constraints to automate complex workflows — from environment setup and data preparation to training launch, monitoring, debugging, and model porting.

Why FlagScale-Agent?

• Domain-Specialized — 23 built-in skills and 13 knowledge domains covering Megatron-LM, TransformerEngine, NCCL, FlashAttention, and more
• Autonomous & Safe — Multi-layer guard system with inject/block/escalate mechanisms prevents runaway execution
• Persistent Intelligence — Cross-session memory system (fact/pitfall/insight) accumulates findings and lessons learned
• Structured Execution — Plan system with acceptance criteria and verification gates ensures quality
• Context-Aware — Automatic context compaction with evict/recall keeps long sessions running smoothly

---

📋 Quick Start

Prerequisites

• Python 3.10 or higher
• API key for Anthropic Claude or OpenAI GPT

Installation

git clone https://github.com/flagos-ai/FlagScale-Agent.git
cd FlagScale-Agent
pip install -e .

Optional extras:

pip install -e ".[checkpoint]"   # enables the inspect_checkpoint tool (pulls torch + safetensors)
pip install -e ".[dev]"          # test/lint tooling (pytest, ruff)

Configuration

Set your API key:

# For Anthropic Claude (recommended)
export ANTHROPIC_API_KEY="your_api_key_here"

# For OpenAI GPT
export OPENAI_API_KEY="your_api_key_here"

Optionally create a config file at ~/.flagscale/agent.yaml:

# LLM Provider
provider: anthropic                        # anthropic | openai
model: claude-sonnet-4-20250514            # Model name
# api_key: sk-xxx                          # Optional: defaults to env var
# base_url: https://api.custom.com/v1      # Optional: custom endpoint

# Execution Limits
max_iterations: 200                        # Max iterations per turn
max_continuations: 200                     # Max consecutive empty responses
max_context_tokens: 0                      # Context window (0 = auto-detect)
max_output_tokens: 8192                    # Max tokens per response

# See examples/agent_config.yaml for full configuration options

Your First Command

# Interactive mode
flagscale-agent

# Single-shot query