AgentForge¶

ReAct agents on open-weight LLMs — tools, memory, and an eval harness. Pairs with ragforge-ml for retrieval and turboquant-ml for quantized model serving.

What it is¶

A small, readable ReAct agent runtime that you can run end-to-end on your own machine with open models, no API key required. Each tool is one short module behind a tiny interface (name, description, run(input) -> str) and every step in the loop is auditable.

Three opinions:

Open models first. Defaults to Qwen/Qwen2.5-3B-Instruct. No OpenAI key required to try.
ReAct, not magic. The loop is a single 60-line function. Compare with any "agent framework" — you can read this one in a coffee break.
Tools with hard limits. Python REPL is AST-whitelisted, SQL is read-only, web search is rate-limited.

Install¶

pip install agentforge-ml                       # core
pip install "agentforge-ml[tools]"              # + sympy + duckduckgo-search
pip install "agentforge-ml[rag]"                # + ragforge-ml
pip install "agentforge-ml[quantized]"          # + turboquant-ml NF4
pip install "agentforge-ml[serve]"              # + FastAPI
pip install "agentforge-ml[all]"                # everything

60-second tour¶

from agentforge import Agent
from agentforge.tools import Calculator, PythonREPL

agent = Agent.from_defaults(
    model_id="Qwen/Qwen2.5-3B-Instruct",
    tools=[Calculator(), PythonREPL()],
)

result = agent.run("What is 47 * 1337, then take its square root?")
print(result.final_answer)
for step in result.steps:
    print(f"  [{step.tool}] {step.action_input!r} -> {step.observation!r}")

CLI¶

af ask "How many primes below 100?" --tools python_repl
af eval examples/eval_set.jsonl
af serve --tools calculator,python_repl

See ReAct, Tools, Memory, Evaluation.