diff --git a/src/backend/services/docker_utils.py b/src/backend/services/docker_utils.py
index 0daf596b35..9a26381f86 100644
--- a/src/backend/services/docker_utils.py
+++ b/src/backend/services/docker_utils.py
@@ -573,11 +573,35 @@ async def containers_run(
     loop = asyncio.get_event_loop()
 
     def _run():
+        _cap_nano_cpus_to_host(kwargs)
         return docker_client.containers.run(image, command=command, **kwargs)
 
     return await loop.run_in_executor(_docker_executor, _run)
 
 
+def _cap_nano_cpus_to_host(kwargs: Dict[str, Any]) -> None:
+    """#2318: Docker rejects NanoCpus above the host's CPU count with a 400, so a
+    template asking for more CPUs than the host has (trinity-system asks for 4)
+    could never be created on a smaller host. Cap the limit here, the one seam
+    every agent create/recreate goes through. The `trinity.cpu` label keeps the
+    requested value: the resource drift check compares it with the DB setting,
+    and a capped label would recreate the agent on every start."""
+    requested = kwargs.get("nano_cpus")
+    if not requested:
+        return
+    try:
+        host_nano = int(docker_client.info()["NCPU"]) * 1_000_000_000
+    except Exception as e:  # noqa: BLE001 — unknown host size: let Docker decide
+        logger.warning(f"[#2318] could not read host CPU count, CPU limit not capped: {e}")
+        return
+    if requested > host_nano:
+        logger.warning(
+            f"[#2318] CPU limit {requested / 1e9:g} exceeds the host's "
+            f"{host_nano / 1e9:g} CPUs; capping to {host_nano / 1e9:g}"
+        )
+        kwargs["nano_cpus"] = host_nano
+
+
 # =============================================================================
 # Container Exec Operations
 # =============================================================================
diff --git a/tests/unit/test_2318_cpu_capped_to_host.py b/tests/unit/test_2318_cpu_capped_to_host.py
new file mode 100644
index 0000000000..6f1257d1c7
--- /dev/null
+++ b/tests/unit/test_2318_cpu_capped_to_host.py
@@ -0,0 +1,59 @@
+"""
+#2318: containers_run caps the requested CPU limit at the Docker host's CPU
+count. Docker rejects NanoCpus above the host's CPUs with a 400, so a template
+asking for cpu "4" (the trinity-system template does) could never be created on
+a 2-CPU host.
+
+Module: src/backend/services/docker_utils.py
+"""
+
+import sys
+from pathlib import Path
+
+import pytest
+
+# Sibling import — the unit dir is not implicitly importable.
+if str(Path(__file__).resolve().parent) not in sys.path:
+    sys.path.insert(0, str(Path(__file__).resolve().parent))
+from test_docker_utils import get_docker_utils  # noqa: E402
+
+GB = 1_000_000_000
+
+
+async def _run_with(host_cpus, **kwargs):
+    docker_utils, client = get_docker_utils()
+    client.info.return_value = {"NCPU": host_cpus}
+    await docker_utils.containers_run("img", detach=True, **kwargs)
+    return client.containers.run.call_args.kwargs
+
+
+@pytest.mark.unit
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "requested,host,expected",
+    [
+        (4, 2, 2),   # the trinity-system template on a t3a.large
+        (4, 4, 4),   # fits: unchanged
+        (2, 8, 2),   # below the host: unchanged
+        (8, 3, 3),   # an odd host count is a valid Docker limit
+    ],
+)
+async def test_nano_cpus_capped_to_host(requested, host, expected):
+    kwargs = await _run_with(host, nano_cpus=requested * GB)
+    assert kwargs["nano_cpus"] == expected * GB
+
+
+@pytest.mark.unit
+@pytest.mark.asyncio
+async def test_no_cpu_limit_is_left_alone():
+    kwargs = await _run_with(2)
+    assert "nano_cpus" not in kwargs
+
+
+@pytest.mark.unit
+@pytest.mark.asyncio
+async def test_unreadable_host_count_passes_request_through():
+    docker_utils, client = get_docker_utils()
+    client.info.side_effect = RuntimeError("daemon unreachable")
+    await docker_utils.containers_run("img", detach=True, nano_cpus=4 * GB)
+    assert client.containers.run.call_args.kwargs["nano_cpus"] == 4 * GB