diff --git a/AGENTS.md b/AGENTS.md
new file mode 100644
index 0000000..e4aeb79
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,16 @@
+# AGENTS.md — benchmark-tool
+
+Follow the org-wide rules in [BenchCommons/.github AGENTS.md](https://github.com/BenchCommons/.github/blob/main/AGENTS.md) and the decisions in [.product](https://github.com/BenchCommons/.product). This file only adds what is specific to this repo.
+
+## Role
+Tool: builds each catalog implementation in Docker, runs it and records reproducible results. It is derived from the TechEmpower toolset. It owns the **catalog contract** and the **published results format** (see `.product/CONTRACTS.md`).
+
+## Stack and commands
+- Python 3.12. `./ssgberk` builds the toolset image and runs `toolset/run-tests.py`.
+- `frameworks/` is a submodule pointing to `BenchCommons/catalog-ssg`.
+- Results are written to `results/<ts>/results.json` and published under `docs/results/<date>/<suite>/<cell>/`.
+- Checks: ruff (py312), pytest, compileall, docker build.
+- Specs and their order: `docs/specs/` (see `ROADMAP.md`).
+
+## Rules
+- Generalizing beyond SSGs (benchmark-tool#29) and switching to perfratio (#30) are tracked in issues. Do not hard-code new SSG-only assumptions.