diff --git a/pycodeloop/core/agent.py b/pycodeloop/core/agent.py
index dd860cb..4f533bb 100644
--- a/pycodeloop/core/agent.py
+++ b/pycodeloop/core/agent.py
@@ -27,6 +27,8 @@
 _MAX_RETRIES = 3
 _RETRY_BASE_DELAY = 1.0
 
+_MAX_EMPTY_RESPONSE_RETRIES = 2
+
 _TOOL_RESULT_SUMMARIZE_THRESHOLD = 8_000
 _TOOL_SUMMARY_PROMPT = (
     "Summarize this tool output concisely, preserving specific facts, "
@@ -415,6 +417,59 @@ def run(
             )
             elapsed = time.perf_counter() - started_at
 
+            empty_retries = 0
+            while (
+                not response.text.strip()
+                and not response.tool_calls
+                and empty_retries < _MAX_EMPTY_RESPONSE_RETRIES
+            ):
+                self.usage = self.usage + response.usage
+                if self.on_usage:
+                    self.on_usage(response.usage, self.usage, elapsed)
+
+                empty_retries += 1
+                self._trace(
+                    "empty_response_retry",
+                    model=self.provider.model,
+                    attempt=empty_retries,
+                )
+                if self.on_retry:
+                    self.on_retry(
+                        empty_retries,
+                        0.0,
+                        RuntimeError(
+                            f"{self.provider.model} returned an empty "
+                            "response with no tool calls"
+                        ),
+                    )
+                started_at = time.perf_counter()
+                response = self._complete(
+                    system_prompt=self.system_prompt,
+                    messages=session.history(),
+                    tools=tools,
+                    on_delta=self.on_text_delta,
+                )
+                elapsed = time.perf_counter() - started_at
+
+            if not response.text.strip() and not response.tool_calls:
+                self.usage = self.usage + response.usage
+                if self.on_usage:
+                    self.on_usage(response.usage, self.usage, elapsed)
+
+                error_text = (
+                    f"{self.provider.model} returned an empty response "
+                    "with no tool calls after "
+                    f"{_MAX_EMPTY_RESPONSE_RETRIES} retries. The model may "
+                    "be too weak for agent mode with the current tool/"
+                    "system-prompt size — try a larger model."
+                )
+                session.add_assistant(error_text)
+                self._notify_message()
+                if self.on_turn_end:
+                    self.on_turn_end()
+                self._trace("run_end", reason="empty_response")
+                return error_text
+
             self.usage = self.usage + response.usage
             if self.on_usage:
                 self.on_usage(response.usage, self.usage, elapsed)
diff --git a/tests/core/test_agent.py b/tests/core/test_agent.py
index e916d53..703202a 100644
--- a/tests/core/test_agent.py
+++ b/tests/core/test_agent.py
@@ -123,6 +123,92 @@ def test_returns_text_when_no_tool_calls(self):
 
         self.assertEqual(result, "hello")
 
+    def test_retries_and_recovers_from_empty_response_with_no_tool_calls(
+        self,
+    ):
+        """Regression: a weak model can return a fully empty message (no
+        text, no tool calls) — the agent used to treat that as a normal
+        `done` turn and silently return "". It should retry instead."""
+        provider = FakeProvider(
+            [
+                ProviderResponse(text="", tool_calls=[]),
+                ProviderResponse(text="", tool_calls=[]),
+                ProviderResponse(text="finally, here's the answer"),
+            ]
+        )
+        agent = Agent(provider=provider, tools=[EchoTool()])
+
+        result = agent.run("hi")
+
+        self.assertEqual(result, "finally, here's the answer")
+
+    def test_gives_up_with_a_clear_error_after_repeated_empty_responses(
+        self,
+    ):
+        provider = FakeProvider(
+            [ProviderResponse(text="", tool_calls=[]) for _ in range(5)]
+        )
+        agent = Agent(provider=provider, tools=[EchoTool()])
+
+        result = agent.run("hi")
+
+        self.assertIn("empty response", result)
+        self.assertIn(provider.model, result)
+
+    def test_accumulates_usage_across_empty_response_retries(self):
+        provider = FakeProvider(
+            [
+                ProviderResponse(
+                    text="", tool_calls=[], usage=Usage(input_tokens=10)
+                ),
+                ProviderResponse(
+                    text="", tool_calls=[], usage=Usage(input_tokens=20)
+                ),
+                ProviderResponse(text="done", usage=Usage(input_tokens=30)),
+            ]
+        )
+        agent = Agent(provider=provider, tools=[EchoTool()])
+
+        agent.run("hi")
+
+        self.assertEqual(agent.usage.input_tokens, 60)
+
+    def test_accumulates_usage_even_when_giving_up_after_empty_responses(
+        self,
+    ):
+        provider = FakeProvider(
+            [
+                ProviderResponse(
+                    text="", tool_calls=[], usage=Usage(input_tokens=5)
+                )
+                for _ in range(3)
+            ]
+        )
+        agent = Agent(provider=provider, tools=[EchoTool()])
+
+        agent.run("hi")
+
+        self.assertEqual(agent.usage.input_tokens, 15)
+
+    def test_records_assistant_turn_in_session_when_giving_up(self):
+        provider = FakeProvider(
+            [ProviderResponse(text="", tool_calls=[]) for _ in range(3)]
+        )
+        notified = []
+        agent = Agent(
+            provider=provider,
+            tools=[EchoTool()],
+            on_message=lambda: notified.append(True),
+        )
+        session = Session(system_prompt="sys")
+
+        result = agent.run("hi", session=session)
+
+        history = session.history()
+        self.assertEqual(history[-1].role, "assistant")
+        self.assertEqual(history[-1].content, result)
+        self.assertTrue(notified)
+
     def test_prefers_explicit_provider_context_window(self):
         provider = FakeProvider(
             [ProviderResponse(text="done", usage=Usage(input_tokens=90))]