A minimum viable agent — ~500 lines of pure Python. No LangChain. No CrewAI. No AutoGen. Just the essential loop, fully readable, endlessly hackable. Understand agents by building the smallest one that actually works.
Every line of code is intentional. No hidden abstractions. If you can read Python, you understand the entire agent.
NanoAgent is a reference implementation. Fork it, break it, extend it. It's the baseline you benchmark everything else against.
Not a toy. NanoAgent will ship with 3 real example agents: a researcher, a coder, and a monitor. Each does something genuinely useful.
Read files, web content, terminal output, structured data
Single-model chain-of-thought. No multi-agent overhead.
Shell, file I/O, web fetch, Python execution, memory read/write
Sliding context window + lightweight vector store for long-term recall
Agent scores its own outputs and retries on failure
Live output so you see the agent's reasoning in real time
Give it a topic. It searches, reads, synthesizes, and writes a structured report.
Give it a task. It writes code, runs tests, fixes errors, and ships working output.
Give it a metric. It watches it, detects anomalies, and alerts you when something's wrong.
NanoAgent is in active development. Join the waitlist for early access and updates.