diff --git a/agent_core/providers/anthropic.py b/agent_core/providers/anthropic.py
index 47e68de..2b8ac6e 100644
--- a/agent_core/providers/anthropic.py
+++ b/agent_core/providers/anthropic.py
@@ -102,6 +102,7 @@ def _build_kwargs(
             for m in msgs
             if (converted := _to_anthropic_msg(m)) is not None
         ]
+        pairs = _merge_tool_results(pairs)
         transient_tail = 0
         for _, is_transient in reversed(pairs):
             if not is_transient:
@@ -378,6 +379,39 @@ async def _prepare_request(self, request: httpx.Request) -> None:
 # ── Conversion helpers ───────────────────────────────────────────────────
 
 
+def _merge_tool_results(
+    pairs: list[tuple[dict[str, Any], bool]],
+) -> list[tuple[dict[str, Any], bool]]:
+    """Fold consecutive tool-result-only user messages into one.
+
+    Each OpenAI ``tool`` message converts to its own user message, so a turn
+    with parallel calls produces several in a row. Anthropic merges them, but
+    the documented shape is ONE user message carrying every ``tool_result``,
+    and translating gateways (llm-hub in front of a non-Claude model) reject
+    the split form: "An assistant message with 'tool_calls' must be followed
+    by tool messages responding to each 'tool_call_id'".
+    """
+    def only_results(msg: dict[str, Any]) -> bool:
+        content = msg.get("content")
+        return (
+            msg.get("role") == "user"
+            and isinstance(content, list)
+            and bool(content)
+            and all(isinstance(b, dict) and b.get("type") == "tool_result" for b in content)
+        )
+
+    out: list[tuple[dict[str, Any], bool]] = []
+    for msg, transient in pairs:
+        if out and not transient and not out[-1][1] and only_results(msg) and only_results(
+            out[-1][0]
+        ):
+            prev = out[-1][0]
+            out[-1] = ({**prev, "content": [*prev["content"], *msg["content"]]}, False)
+        else:
+            out.append((msg, transient))
+    return out
+
+
 def _split_system(messages: list[Message]) -> tuple[str, list[Message]]:
     """Pull out the (single) leading system message; Anthropic takes it
     as a top-level kwarg, not as a message."""
diff --git a/changes/anthropic-merge-tool-results.fix.md b/changes/anthropic-merge-tool-results.fix.md
new file mode 100644
index 0000000..b34710f
--- /dev/null
+++ b/changes/anthropic-merge-tool-results.fix.md
@@ -0,0 +1 @@
+The Anthropic client now sends the tool results of one turn's parallel tool calls in a single user message, as Anthropic documents, instead of one user message per result. Translating gateways (llm-hub in front of a non-Claude model such as deepseek-flash) rejected the split form with a 400.
diff --git a/tests/test_provider_native_clients.py b/tests/test_provider_native_clients.py
index 515aa9d..6d1726d 100644
--- a/tests/test_provider_native_clients.py
+++ b/tests/test_provider_native_clients.py
@@ -1190,3 +1190,34 @@ def test_transient_flag_is_not_sent_on_openai_wire():
     from agent_core.messages import for_wire
 
     assert for_wire([_transient("x")]) == [user_msg("x")]
+
+
+def test_anthropic_parallel_tool_results_share_one_user_message(monkeypatch):
+    """Translating gateways reject one user message per tool_result."""
+    monkeypatch.setenv("ANTHROPIC_PROMPT_CACHE", "0")
+    c = ac.AnthropicClient("claude-x", api_key="x")
+    calls = [{"id": i, "type": "function", "function": {"name": "bash", "arguments": "{}"}}
+             for i in ("a", "b")]
+    kwargs = c._build_kwargs(
+        [system_msg("s"), user_msg("q"), assistant_msg("", tool_calls=calls),
+         tool_msg("ra", "a"), tool_msg("rb", "b"), _transient("[env]")],
+        tools=None, temperature=None, max_tokens=None, extra_headers=None, timeout=None,
+    )
+    msgs = kwargs["messages"]
+    assert [m["role"] for m in msgs] == ["user", "assistant", "user", "user"]
+    assert [b["tool_use_id"] for b in msgs[2]["content"]] == ["a", "b"]
+    assert msgs[3]["content"] == "[env]"
+
+
+def test_anthropic_cache_breakpoint_after_merged_tool_results(monkeypatch):
+    monkeypatch.delenv("ANTHROPIC_PROMPT_CACHE", raising=False)
+    c = ac.AnthropicClient("claude-x", api_key="x")
+    calls = [{"id": i, "type": "function", "function": {"name": "bash", "arguments": "{}"}}
+             for i in ("a", "b")]
+    kwargs = c._build_kwargs(
+        [system_msg("s"), user_msg("q"), assistant_msg("", tool_calls=calls),
+         tool_msg("ra", "a"), tool_msg("rb", "b"), _transient("[env]")],
+        tools=None, temperature=None, max_tokens=None, extra_headers=None, timeout=None,
+    )
+    assert _tail_breakpoints(kwargs) == [2]
+    assert "cache_control" in kwargs["messages"][2]["content"][-1]