diff --git a/src/agents/sandbox/session/pty_output.py b/src/agents/sandbox/session/pty_output.py
index 25cbe774e7..84ac46a762 100644
--- a/src/agents/sandbox/session/pty_output.py
+++ b/src/agents/sandbox/session/pty_output.py
@@ -8,6 +8,33 @@
 from .pty_types import truncate_text_by_tokens
 
 
+def _incomplete_utf8_suffix_length(data: bytes | bytearray) -> int:
+    """Return how many trailing bytes start a UTF-8 sequence that is not finished yet."""
+    # A sequence is at most four bytes, so the lead byte is within four of the end.
+    for back in range(1, min(4, len(data)) + 1):
+        byte = data[-back]
+        if byte < 0x80:
+            # ASCII cannot be part of a multi-byte sequence, so nothing is pending.
+            return 0
+        if byte < 0xC0:
+            # Continuation byte. Keep walking back to find the lead byte it belongs to.
+            continue
+        # Only real lead bytes can still be completed. 0xC0, 0xC1 and 0xF5 to 0xFF never
+        # start a valid sequence, so carrying them would withhold a byte that no later
+        # output can finish and would suppress the replacement character forever.
+        if 0xC2 <= byte <= 0xDF:
+            needed = 2
+        elif 0xE0 <= byte <= 0xEF:
+            needed = 3
+        elif 0xF0 <= byte <= 0xF4:
+            needed = 4
+        else:
+            return 0
+        return back if back < needed else 0
+    # Four trailing bytes with no lead byte cannot be completed either.
+    return 0
+
+
 async def collect_pty_output(
     *,
     output_chunks: deque[bytes],
@@ -45,6 +72,18 @@ async def collect_pty_output(
             break
         output_notify.clear()
 
+    if not is_done():
+        # A multi-byte character can straddle two collection windows. Decoding a partial
+        # sequence with errors="replace" destroys those bytes, so hold the unfinished tail
+        # back for the next window. Once the producer is done nothing can complete it, so
+        # the replacement behaviour below is the right answer then.
+        carry = _incomplete_utf8_suffix_length(output)
+        if carry:
+            tail = bytes(output[-carry:])
+            del output[-carry:]
+            async with output_lock:
+                output_chunks.appendleft(tail)
+
     text = output.decode("utf-8", errors="replace")
     truncated, original_token_count = truncate_text_by_tokens(text, max_output_tokens)
     return truncated.encode("utf-8", errors="replace"), original_token_count
diff --git a/tests/sandbox/test_pty_output.py b/tests/sandbox/test_pty_output.py
index f15bcf85b4..cb06330773 100644
--- a/tests/sandbox/test_pty_output.py
+++ b/tests/sandbox/test_pty_output.py
@@ -5,7 +5,10 @@
 
 import pytest
 
-from agents.sandbox.session.pty_output import collect_pty_output
+from agents.sandbox.session.pty_output import (
+    _incomplete_utf8_suffix_length,
+    collect_pty_output,
+)
 
 
 @pytest.mark.asyncio
@@ -57,3 +60,146 @@ def mark_done() -> bool:
 
     assert output == b"before done after done"
     assert original_token_count is None
+
+
+@pytest.mark.asyncio
+async def test_collect_pty_output_holds_split_utf8_character_for_the_next_window() -> None:
+    """A character split across two collection windows must survive, not become U+FFFD."""
+    text = "h\u00e9llo w\u00f6rld \u2705"
+    raw = text.encode("utf-8")
+    split = 2  # after "h" and the first byte of the two-byte "e" with acute
+
+    output_chunks: deque[bytes] = deque()
+    output_lock = asyncio.Lock()
+    output_notify = asyncio.Event()
+    producer_done = False
+
+    async def window() -> bytes:
+        output_notify.set()
+        collected, _ = await collect_pty_output(
+            output_chunks=output_chunks,
+            output_lock=output_lock,
+            output_notify=output_notify,
+            is_done=lambda: producer_done,
+            yield_time_ms=1,
+            max_output_tokens=None,
+        )
+        return collected
+
+    output_chunks.append(raw[:split])
+    first = await window()
+    # The partial sequence is withheld rather than decoded, so the window ends on "h".
+    assert first == b"h"
+
+    output_chunks.append(raw[split:])
+    producer_done = True
+    second = await window()
+
+    assert (first + second).decode("utf-8") == text
+
+
+@pytest.mark.asyncio
+async def test_collect_pty_output_replaces_partial_utf8_once_the_producer_is_done() -> None:
+    """Nothing can complete a truncated sequence after the producer closes, so replace it."""
+    output_chunks: deque[bytes] = deque([b"ok\xc3"])
+    output_notify = asyncio.Event()
+    output_notify.set()
+
+    output, _ = await collect_pty_output(
+        output_chunks=output_chunks,
+        output_lock=asyncio.Lock(),
+        output_notify=output_notify,
+        is_done=lambda: True,
+        yield_time_ms=1,
+        max_output_tokens=None,
+    )
+
+    assert output.decode("utf-8") == "ok\ufffd"
+    assert not output_chunks
+
+
+@pytest.mark.parametrize("lead", [b"\xc0", b"\xc1", b"\xf5", b"\xff"])
+@pytest.mark.asyncio
+async def test_collect_pty_output_replaces_bytes_that_cannot_start_a_character(
+    lead: bytes,
+) -> None:
+    """Only real lead bytes may be carried; others can never be completed by later output."""
+    output_chunks: deque[bytes] = deque([b"prompt" + lead])
+    output_notify = asyncio.Event()
+    output_notify.set()
+
+    output, _ = await collect_pty_output(
+        output_chunks=output_chunks,
+        output_lock=asyncio.Lock(),
+        output_notify=output_notify,
+        is_done=lambda: False,
+        yield_time_ms=1,
+        max_output_tokens=None,
+    )
+
+    # Replaced in this window rather than withheld from an interactive prompt forever.
+    assert output.decode("utf-8") == "prompt\ufffd"
+    assert not output_chunks
+
+
+@pytest.mark.parametrize(
+    ("data", "expected"),
+    [
+        (b"abc", 0),
+        (b"a\xc3", 1),
+        (b"a\xc3\xa9", 0),
+        (b"a\xe2", 1),
+        (b"a\xe2\x82", 2),
+        (b"a\xe2\x82\xac", 0),
+        (b"a\xf0", 1),
+        (b"a\xf0\x9f", 2),
+        (b"a\xf0\x9f\x98", 3),
+        (b"a\xf0\x9f\x98\x80", 0),
+        (b"a\xc0", 0),
+        (b"a\xc1", 0),
+        (b"a\xf5", 0),
+        (b"a\xff", 0),
+        (b"a\x80", 0),
+        (b"\x80\x80\x80\x80", 0),
+    ],
+)
+def test_incomplete_utf8_suffix_length(data: bytes, expected: int) -> None:
+    """Every split position of every sequence length, plus bytes that can never complete."""
+    assert _incomplete_utf8_suffix_length(data) == expected
+
+
+@pytest.mark.parametrize("split", [1, 2])
+@pytest.mark.asyncio
+async def test_collect_pty_output_holds_three_byte_character_split_at_any_point(
+    split: int,
+) -> None:
+    """A three-byte character must survive a window boundary after one or two bytes."""
+    text = "a\u20acb"
+    raw = text.encode("utf-8")
+    boundary = 1 + split  # after "a" plus part of the euro sign
+
+    output_chunks: deque[bytes] = deque([raw[:boundary]])
+    output_lock = asyncio.Lock()
+    output_notify = asyncio.Event()
+    producer_done = False
+
+    async def window() -> bytes:
+        output_notify.set()
+        collected, _ = await collect_pty_output(
+            output_chunks=output_chunks,
+            output_lock=output_lock,
+            output_notify=output_notify,
+            is_done=lambda: producer_done,
+            yield_time_ms=1,
+            max_output_tokens=None,
+        )
+        return collected
+
+    first = await window()
+    assert first == b"a"
+
+    output_chunks.append(raw[boundary:])
+    producer_done = True
+    second = await window()
+
+    assert (first + second).decode("utf-8") == text