diff --git a/.github/workflows/commit_checks.yml b/.github/workflows/commit_checks.yml
index 41300120..494c422a 100644
--- a/.github/workflows/commit_checks.yml
+++ b/.github/workflows/commit_checks.yml
@@ -48,6 +48,19 @@ jobs:
       - name: Run pytest
         run: pytest -v -s -ra --tb=long -o log_cli=true -o log_level=DEBUG
 
+  run-benchmarks:
+    name: Smoke Test Benchmarks
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v4
+      - uses: ./.github/actions/build
+        with:
+          python-version: 3.11
+          pip-options: all,rllib
+      - name: Smoke test benchmark envs
+        run: python smoke_test.py
+        working-directory: benchmarks
+
   check-doc-build:
     name: Test Documentation Build
     runs-on: ubuntu-latest
diff --git a/benchmarks/smoke_test.py b/benchmarks/smoke_test.py
new file mode 100644
index 00000000..0f45b0e0
--- /dev/null
+++ b/benchmarks/smoke_test.py
@@ -0,0 +1,76 @@
+"""Smoke-test the benchmark environments.
+
+This constructs each benchmark environment from its ``BenchmarkEnv.env_args``
+and steps it a few times to verify that it builds and runs without error. It
+does not train a policy; it only exercises environment construction and the
+Gymnasium/PettingZoo step loop so that CI can catch breakages in the benchmark
+definitions quickly. Full training is driven by ``benchmark.py``.
+
+Run from within the ``benchmarks`` directory so the benchmark modules (which
+use top-level imports like ``from benchmark import BenchmarkEnv``) are
+importable::
+
+    cd benchmarks
+    python smoke_test.py
+"""
+
+import argparse
+import sys
+from copy import deepcopy
+
+from aeos import aeos_single
+from nadir_science import nadir_science
+from rso_inspection import rso_inspection
+
+from bsk_rl import ConstellationTasking
+
+# One representative environment per benchmark module. Add entries here to
+# smoke-test additional benchmark environments.
+BENCHMARK_ENVS = {
+    "nadir_science": nadir_science,
+    "aeos_single": aeos_single,
+    "rso_inspection": rso_inspection,
+}
+
+
+def smoke_test_env(name, benchmark_env, n_steps):
+    """Build a benchmark environment and step it ``n_steps`` times."""
+    print(f"Smoke-testing benchmark env '{name}'...")
+
+    # env_args carries RLlib-only callbacks that ConstellationTasking does not
+    # accept. Strip them the same way the RLlib env creator does.
+    env_args = deepcopy(benchmark_env.env_args)
+    env_args.pop("episode_data_callback", None)
+    env_args.pop("satellite_data_callback", None)
+
+    env = ConstellationTasking(**env_args)
+    env.reset(seed=0)
+    for step in range(n_steps):
+        action = {agent: env.action_space(agent).sample() for agent in env.agents}
+        _, _, terminated, truncated, _ = env.step(action)
+        if all(terminated.values()) or all(truncated.values()):
+            break
+    env.close()
+    print(f"Benchmark env '{name}' completed {step + 1} step(s).")
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "-n",
+        "--n_steps",
+        type=int,
+        default=3,
+        help="Number of environment steps to run per benchmark env.",
+    )
+    args = parser.parse_args()
+
+    for name, benchmark_env in BENCHMARK_ENVS.items():
+        smoke_test_env(name, benchmark_env, args.n_steps)
+
+    print(f"All {len(BENCHMARK_ENVS)} benchmark env(s) smoke-tested successfully.")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())