loop_engineering_lab

Ebook: https://shop.beacons.ai/aiengineeringinsider/63560a2a-7b0b-4869-bf8f-5250b19ae8a6

Preview: https://drive.google.com/file/d/1twfTGCgw69mLccQcyADrve9LmXYG3z35/view?usp=sharing

A small, runnable reference implementation that accompanies Loop Engineering: A Practitioner's Guide. Each module maps to a chapter so you can read the book and the source side by side.

It runs with no API key by default — the LLM is a deterministic mock, so the demos and tests are reproducible offline and in CI.

Layout

Path	Chapter	Purpose
llm/base.py	2	LLMClient protocol, Message, LLMResponse
llm/mock.py	2	Deterministic, scripted mock model
llm/openai_adapter.py	2	Optional real provider (no hard dependency)
core/state.py	2	Loop state machine (LoopPhase, LoopState)
core/control.py	3	Budget, TerminationPolicy, stopping rules
core/agent_loop.py	2	The observe–reason–act loop
tools/	6	Tool registry + sandboxed built-ins
memory/	4	Working + episodic memory
verifiers/	8	Verdict-returning verifiers
observability/	13	Structured tracer
tests/	—	Hermetic unit tests

Run

# from the repository root
python -m loop_engineering_lab demo      # a verified ReAct loop, end to end
python -m loop_engineering_lab budget     # a runaway loop stopped by its budget

# tests
pip install pytest
pytest loop_engineering_lab/tests

Use a real model (optional)

pip install openai
export OPENAI_API_KEY=sk-...
python -m loop_engineering_lab demo --real

The agent loop depends only on the LLMClient protocol, so swapping the mock for a real provider changes no orchestration code — that seam is the point.

loop-engineering