multi-agent-scm-bench

A 4-tier agentic supply chain testbed (retailer → wholesaler → distributor → factory, i.e. a classical beer game setup). Lower cost wins.

Why

In SCM, ten strands shift across three eras:

Strand	Era I: Networks & Relationships	Era II: Platforms & Ecosystems	Era III: Protocols & Agents
Structure	Networks	Platforms / Ecosystems	Protocols
Execution	Machine	Automation	Autonomy
Power	Control	Co-creation	Agency
Methods	Production	Orchestration	Co-design
Technology	Force	Information	Intelligence
Transformations	Processes	Flows	Evolution
Dominant logic	Goods (GDL)	Service (SDL)	Agency (ADL)
Anchor actors	Replicators	Mediators	Explorers
Comparative advantage	Factor endowments	Institutional arrangements	Programmable capacity
Value	Value-in-Exchange	Value-in-Use	Survival

Right now, we lack benchmarking and simulation for these shifts. This bench offers a starting point with protocols, autonomy, and agency. Future releases may focus on agent co-design, the value of information as the fundamental unit of exchange, agency levels, the impact of intelligence (e.g. different models, modalities), and the role of evolution as survival in an intelligence explosion scenario. Further, adversarial elements require attention.

Layout

• src/scm_bench/ — engine, validator, runner, metrics, LLM runtime
• examples/mirror/ — baseline agent with a deterministic policy (proportional-to-incoming, composite 1.000)
• examples/llm/ — LLM-driven agent via transformers / vLLM / OpenAI-compatible (including local Ollama)
• tests/ — pytest suite

Bundle contracts and run/scoring instructions live in each examples/<name>/AGENTS.md.

pip install -e .
pytest

---

Tools and hardware: ATTRIBUTION.md. Dependencies: pyproject.toml. License: 0BSD (LICENSE).