diff --git a/tests/test_get_workers.py b/tests/test_get_workers.py
new file mode 100644
index 0000000..558c528
--- /dev/null
+++ b/tests/test_get_workers.py
@@ -0,0 +1,143 @@
+"""Smoke tests for get_workers: provider diversity, tier filtering, degraded handling."""
+
+import json
+from unittest.mock import AsyncMock, patch
+
+import pytest
+
+from model_radar.providers import Model
+from model_radar.scanner import PingResult
+
+
+def _model(model_id: str, provider: str, tier: str = "S") -> Model:
+    """Create a test model using a real provider key."""
+    return Model(
+        model_id=model_id,
+        label=model_id,
+        tier=tier,
+        swe_score="60%",
+        context="128k",
+        provider=provider,
+    )
+
+
+def _ping(model: Model, latency_ms: float = 200.0) -> PingResult:
+    return PingResult(model=model, status="up", latency_ms=latency_ms)
+
+
+# Use real provider keys so format_result can look up PROVIDERS[key].name
+P1, P2, P3, P4, P5 = "nvidia", "groq", "cerebras", "openrouter", "sambanova"
+
+
+@pytest.mark.asyncio
+async def test_get_workers_returns_provider_diverse_results():
+    """get_workers returns exactly N models from N distinct providers."""
+    from model_radar.server import get_workers
+
+    models = [
+        _ping(_model("model-a", P1, "S+"), 100),
+        _ping(_model("model-a2", P1, "S"), 110),  # same provider, should be skipped
+        _ping(_model("model-b", P2, "S"), 150),
+        _ping(_model("model-c", P3, "A+"), 200),
+        _ping(_model("model-d", P4, "A"), 250),
+        _ping(_model("model-e", P5, "A"), 300),
+    ]
+
+    with patch("model_radar.server.scan_models", new_callable=AsyncMock, return_value=models):
+        result = json.loads(await get_workers(count=4, min_tier="A", verified=False))
+
+    assert result["count"] == 4
+    assert result["distinct_providers"] == 4
+
+    providers = [w["provider_key"] for w in result["workers"]]
+    assert len(set(providers)) == 4, f"Expected 4 distinct providers, got {providers}"
+
+    # Should pick one from each of the first 4 distinct providers (skipping model-a2)
+    model_ids = [w["model_id"] for w in result["workers"]]
+    assert model_ids == ["model-a", "model-b", "model-c", "model-d"]
+
+
+@pytest.mark.asyncio
+async def test_get_workers_respects_tier_filter():
+    """All returned workers must meet the min_tier threshold."""
+    from model_radar.server import get_workers
+
+    # scan_models handles tier filtering upstream, so simulate pre-filtered results
+    filtered = [
+        _ping(_model("s-plus", P1, "S+"), 100),
+        _ping(_model("s-tier", P2, "S"), 150),
+        _ping(_model("a-tier", P3, "A"), 200),
+    ]
+
+    with patch("model_radar.server.scan_models", new_callable=AsyncMock, return_value=filtered):
+        result = json.loads(await get_workers(count=4, min_tier="A", verified=False))
+
+    for w in result["workers"]:
+        assert w["tier"] in {"S+", "S", "A+", "A", "A-"}, f"Unexpected tier {w['tier']}"
+
+
+@pytest.mark.asyncio
+async def test_get_workers_skips_down_models():
+    """Only UP models are included — down/error models are filtered out."""
+    from model_radar.server import get_workers
+
+    models = [
+        _ping(_model("up-1", P1), 100),
+        PingResult(model=_model("down-1", P2), status="error", latency_ms=None),
+        PingResult(model=_model("timeout-1", P3), status="timeout", latency_ms=None),
+        _ping(_model("up-2", P4), 200),
+    ]
+
+    with patch("model_radar.server.scan_models", new_callable=AsyncMock, return_value=models):
+        result = json.loads(await get_workers(count=4, min_tier="A", verified=False))
+
+    assert result["count"] == 2
+    model_ids = [w["model_id"] for w in result["workers"]]
+    assert model_ids == ["up-1", "up-2"]
+
+
+@pytest.mark.asyncio
+async def test_get_workers_graceful_when_fewer_providers_available():
+    """When fewer providers are available than requested, returns what it can."""
+    from model_radar.server import get_workers
+
+    models = [
+        _ping(_model("only-1", P1), 100),
+        _ping(_model("only-2", P2), 200),
+    ]
+
+    with patch("model_radar.server.scan_models", new_callable=AsyncMock, return_value=models):
+        result = json.loads(await get_workers(count=5, min_tier="A", verified=False))
+
+    assert result["count"] == 2
+    assert result["distinct_providers"] == 2
+
+
+@pytest.mark.asyncio
+async def test_get_workers_skips_degraded_then_relaxes():
+    """Degraded providers are skipped initially but included if needed to fill count."""
+    from model_radar.server import get_workers, _state
+
+    models = [
+        _ping(_model("fast", P1), 100),
+        _ping(_model("degraded-model", P2), 150),
+        _ping(_model("slow", P3), 300),
+    ]
+
+    # Mark P2 as degraded (need >=2 recent 429s)
+    _state.throttle.record_429(P2)
+    _state.throttle.record_429(P2)
+
+    try:
+        with patch("model_radar.server.scan_models", new_callable=AsyncMock, return_value=models):
+            result = json.loads(await get_workers(count=3, min_tier="A", verified=False))
+
+        assert result["count"] == 3
+        # All 3 providers used (degraded one included in relaxation pass to fill count)
+        providers = {w["provider_key"] for w in result["workers"]}
+        assert len(providers) == 3
+    finally:
+        # Clean up throttle state
+        _state.throttle._recent_429s.pop(P2, None)
+        _state.throttle._recent_calls.pop(P2, None)
+        _state.throttle._concurrency.pop(P2, None)