Agent Forge

Multi-agent LLM orchestration for Python — Planner → Executor → Critic, with tool-calling, evals, observability, and a keyless LLM backend.

View on GitHub ★ Quick start

What it does

A task goes in; a Planner decomposes it into steps, an Executor runs each one (calling tools where needed), and a Critic judges the result. Ships an LLM-as-judge eval harness, OpenTelemetry spans, and a keyless provider chain — build and benchmark agents with no API key.

Highlights

Planner / Executor / Critic pipeline
Pluggable tools via a @tool decorator
JSON Schema arg validation

Keyless LLM (kilo.ai → pollinations.ai → g4f)
Eval harness with llm_judge assertions
structlog + OpenTelemetry observability

Install & run

git clone https://github.com/chirag127/agent-forge
cd agent-forge
pip install -e .

agent-forge run "What is the square root of 2025?"
agent-forge eval

Requires Python 3.13. PyPI publish is on the roadmap.