diff --git a/pyproject.toml b/pyproject.toml
index 69e9694e..2fde6dac 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -30,8 +30,10 @@ jaxcpu = [
 dev = [
     "build",
     "flake8",
+    "google-genai",
     "isort",
     "mypy>=1.0.0",
+    "pydantic",
     "pytest",
     "regex",
     "toml",
@@ -59,8 +61,10 @@ budoux = ["models/*.json", "skip_nodes.json", "py.typed"]
 dev = [
     "build",
     "flake8",
+    "google-genai",
     "isort",
     "mypy>=1.0.0",
+    "pydantic",
     "pytest",
     "regex",
     "toml",
diff --git a/scripts/synthesize_samples.py b/scripts/synthesize_samples.py
new file mode 100644
index 00000000..83b4d1b3
--- /dev/null
+++ b/scripts/synthesize_samples.py
@@ -0,0 +1,381 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Agentic candidate sentence synthesis and pre-filtering utility.
+
+Implements a compact 3-agent structure dedicated to synthesizing natural
+training sentences and repairing segmentation discrepancies:
+  1. Intent & Repro Agent: Extracts target break boundaries and confirms live
+     bug reproduction against current BudouX models.
+  2. Example Generator Agent: Synthesizes high-density realistic sentences
+     containing the targeted phrases or break expressions across any language.
+  3. Base Parser Overlay & Linguistic Expert Agent: Maps surrounding sentence
+     breaks strictly to existing BudouX model behavior while accurately pruning
+     awkward linguistic expressions.
+"""
+
+import argparse
+import json
+import os
+import re
+import sys
+from typing import Any, List, Optional
+
+# Module hack to allow importing budoux from repository root when run from CLI
+LIB_PATH = os.path.join(os.path.dirname(__file__), '..')
+sys.path.insert(0, os.path.abspath(LIB_PATH))
+
+import requests  # noqa: E402
+from google import genai  # noqa: E402
+from pydantic import BaseModel, Field  # noqa: E402
+
+import budoux  # noqa: E402
+
+SEP = budoux.utils.SEP
+
+# ==============================================================================
+# Structured Pydantic Models (Contractual Type Boundaries)
+# ==============================================================================
+
+
+class IntentTarget(BaseModel):
+  """Target bug remediation boundaries identified by Intent Understanding."""
+  target_phrase: str = Field(..., description="Raw target phrase string.")
+  expected_split: str = Field(...,
+                              description="Expected string separated with ▁.")
+  is_positive: bool = Field(
+      ...,
+      description="True if forced break across slash, False if cohesive word.")
+  is_reproducible: bool = Field(
+      default=True,
+      description="Whether current runtime model exhibits the reported bug.")
+
+
+class CandidateBatch(BaseModel):
+  """Container for oversampled candidate sentences generated by the Example Generator Agent."""
+  candidates: List[str]
+
+
+class ExpertEvaluation(BaseModel):
+  """Evaluation assessment and local boundary polish for a candidate sentence."""
+  sentence: str = Field(...,
+                        description="Candidate sentence string under audit.")
+  is_fluent: bool = Field(
+      ..., description="True if candidate can be polished/retained naturally.")
+  corrected_sentence: str = Field(
+      "",
+      description="The polished candidate string with proper native boundary splits applied."
+  )
+  rejection_reason: str = Field("", description="Explanation if rejected.")
+
+
+class EvaluationBatch(BaseModel):
+  """Container for linguistic expert audit decisions."""
+  evaluations: List[ExpertEvaluation]
+
+
+# ==============================================================================
+# Agent Utilities & Transformations
+# ==============================================================================
+
+
+def parse_direct_input(input_str: str) -> IntentTarget:
+  """Parses a direct --input parameter into an IntentTarget representation."""
+  normalized = input_str.strip().replace(" / ", SEP).replace("/", SEP)
+  is_positive = SEP in normalized
+  raw = normalized.replace(SEP, "")
+  return IntentTarget(
+      target_phrase=raw,
+      expected_split=normalized,
+      is_positive=is_positive,
+      is_reproducible=True,
+  )
+
+
+def fetch_github_issue_context(issue_id: str,
+                               requests_mod: Any = requests) -> str:
+  """Safely retrieves public GitHub issue description content."""
+  issue_num = re.sub(r"[^\d]", "", str(issue_id))
+  if not issue_num:
+    raise ValueError(f"Invalid issue parameter value: {issue_id}")
+  url = f"https://api.github.com/repos/google/budoux/issues/{issue_num}"
+  response = requests_mod.get(url, timeout=10)
+  response.raise_for_status()
+  data = response.json()
+  raw_body = str(data.get("body") or data.get("title") or "")
+  return raw_body.strip()
+
+
+def verify_bug_reproduction(target: IntentTarget,
+                            parser: budoux.Parser) -> bool:
+  """Tests whether the live parser currently reproduces the reported breaking defect.
+
+  Returns True if the current model segments the string improperly (bug confirmed).
+  Returns False if the current model already achieves expected_split.
+  """
+  clean_text = target.target_phrase.replace(SEP, "")
+  current_splits = parser.parse(clean_text)
+  actual_output = SEP.join(current_splits)
+  return actual_output != target.expected_split
+
+
+def align_to_base_parser_splits(
+    candidates: List[str],
+    target: IntentTarget,
+    parser: budoux.Parser,
+) -> List[str]:
+  """Aligns candidate boundaries outside the target phrase to base parser splits.
+
+  Args:
+    candidates: List of raw candidate sentence strings (`List[str]`).
+    target: IntentTarget parameter declaring `target_phrase` and `expected_split`.
+    parser: Loaded `budoux.Parser` instance (`ja.json`).
+
+  Returns:
+    List of valid candidate strings aligned to base model segmentations outside the target interval.
+  """
+  surviving: List[str] = []
+  raw_target = target.target_phrase.replace(SEP, "")
+  if not raw_target:
+    return surviving
+
+  # Build the regex pattern once outside the loop to locate the target phrase in the parsed string
+  char_patterns = [re.escape(char) for char in raw_target]
+  target_pattern = re.compile(f"[{SEP}]*".join(char_patterns))
+
+  for sentence in candidates:
+    if target.expected_split not in sentence:
+      continue  # Filter: Discard if the targeted split sequence is missing
+
+    # Align: Parse sentence with base model to get default production splits
+    clean_sentence = sentence.replace(SEP, "")
+    base_parsed = SEP.join(parser.parse(clean_sentence))
+
+    # Overlay: Replace the target phrase region in base splits with the corrected split
+    aligned_sentence = target_pattern.sub(
+        lambda m: target.expected_split, base_parsed, count=1)
+    surviving.append(aligned_sentence)
+
+  return surviving
+
+
+# ==============================================================================
+# Agent Roles
+# ==============================================================================
+
+
+def extract_intent_target(
+    issue_text: str,
+    client: genai.Client,
+    model_name: str = "gemini-3.1-flash-lite",
+) -> IntentTarget:
+  """Intent Understanding Agent: diagnoses target bug parameters from issue text."""
+  prompt = f"""You are an expert BudouX segmentation analyst.
+Examine this GitHub bug issue description:
+---
+{issue_text}
+---
+Determine the primary problematic phrase and its desired canonical split formatted with '▁'.
+If a phrase was mistakenly split internally and should stay unbroken (e.g., 'もはや'), return target_phrase='もはや', expected_split='もはや', is_positive=False.
+If a phrase needed a boundary across separate words that was missing or shifted (e.g., 'いよいよ/はじまる'), format expected_split inserting '▁' at the intended break location and set is_positive=True.
+"""
+  config = genai.types.GenerateContentConfig(
+      response_mime_type="application/json",
+      response_schema=IntentTarget,
+      temperature=0.1,
+  )
+  resp = client.models.generate_content(
+      model=model_name, contents=prompt, config=config)
+  return IntentTarget.model_validate_json(resp.text or "")
+
+
+def generate_oversample_candidates(
+    target: IntentTarget,
+    client: genai.Client,
+    num_candidates: int = 30,
+    lang: str = "ja",
+    model_name: str = "gemini-3.1-flash-lite",
+) -> List[str]:
+  """Example Generator Agent: synthesizes diverse natural sentences cleanly across target lang."""
+  prompt = f"""You are a skilled corpus synthesizer for target language code '{lang}'.
+Generate exactly {num_candidates} highly natural, realistic sentences across language '{lang}' where the phrase '{target.expected_split}' appears cleanly within distinct realistic contexts (e.g., news headlines, casual speech, essays, formal letters).
+
+CRITICAL CONSTRAINTS:
+1. Every candidate MUST contain the exact substring '{target.expected_split}' strictly as formatted.
+2. Maintain natural grammatical structures and standard Bunsetsu word boundaries across language '{lang}'.
+"""
+  config = genai.types.GenerateContentConfig(
+      response_mime_type="application/json",
+      response_schema=CandidateBatch,
+      temperature=0.4,
+  )
+  resp = client.models.generate_content(
+      model=model_name, contents=prompt, config=config)
+  batch = CandidateBatch.model_validate_json(resp.text or "")
+  return batch.candidates
+
+
+def prune_linguistic_anomalies(
+    candidates: List[str],
+    client: genai.Client,
+    max_keep: int = 15,
+    lang: str = "ja",
+    model_name: str = "gemini-3.1-flash-lite",
+) -> List[str]:
+  """Linguistic Expert Agent: polishes boundary alignment across candidates and audits fluency."""
+  if not candidates:
+    return []
+
+  lines_summary = [f"{i}. {s}" for i, s in enumerate(candidates)]
+  prompt = f"""You are an expert linguistic polisher and boundary arbiter for language '{lang}'.
+Review these candidate training sentences containing existing boundary splits marked by '▁':
+{chr(10).join(lines_summary)}
+
+Your critical mission across three synthesized layers (existing base parser behavior, intent understanding, and linguistic fluency):
+1. Evaluate whether the sentence is natural and well-structured across target language '{lang}' (is_fluent). Immediately reject (is_fluent=False) malformed syntax or broken words.
+2. For fluent sentences (is_fluent=True), perform precise boundary polishing and return the perfected string in 'corrected_sentence':
+   - Keep quotation framing clean without forcing splits directly after opening brackets or before closing quotes (e.g., maintain '「こんにちは」と▁...' without cutting around quotation marks).
+   - If an independent vocabulary word or bunsetsu was improperly fused due to base model deficiencies (e.g., '彼の▁才能はもはや誰にも▁...' where words like 'もはや' lack surrounding breaks), restore clean natural bunsetsu gaps around the word ('彼の▁才能は▁もはや▁誰にも▁...').
+   - Never alter or remove the canonical splits established inside the targeted phrase itself.
+"""
+  config = genai.types.GenerateContentConfig(
+      response_mime_type="application/json",
+      response_schema=EvaluationBatch,
+      temperature=0.1,
+  )
+  resp = client.models.generate_content(
+      model=model_name, contents=prompt, config=config)
+  eval_batch = EvaluationBatch.model_validate_json(resp.text or "")
+
+  kept: List[str] = []
+  for e in eval_batch.evaluations:
+    if e.is_fluent:
+      kept.append(e.corrected_sentence.strip() or e.sentence)
+  if not kept:
+    kept = candidates
+  return kept[:max_keep]
+
+
+# ==============================================================================
+# Executive Driver & CLI Setup
+# ==============================================================================
+
+
+def run_agentic_synthesis_pipeline(
+    input_str: Optional[str] = None,
+    issue_id: Optional[str] = None,
+    num_candidates: int = 30,
+    max_keep: int = 15,
+    lang: str = "ja",
+    outfile: str = "staging_raw.txt",
+    client: Optional[genai.Client] = None,
+    parser: Optional[budoux.Parser] = None,
+) -> List[str]:
+  """Orchestrates Intent Understanding, Oversample Generation, and Expert Pruning."""
+  if not parser:
+    model_path = os.path.join(
+        os.path.dirname(__file__), "..", "budoux", "models", f"{lang}.json")
+    with open(model_path, "r", encoding="utf-8") as f:
+      parser = budoux.Parser(json.load(f))
+
+  if not client and (issue_id or num_candidates > 0):
+    api_key = os.environ.get("GEMINI_API_KEY", "").strip()
+    if not api_key:
+      raise RuntimeError(
+          "GEMINI_API_KEY environment variable not set or empty.")
+    client = genai.Client(api_key=api_key)
+
+  # Step 1: Intent & Bug Verification
+  if input_str:
+    target = parse_direct_input(input_str)
+  elif issue_id:
+    assert client is not None
+    context = fetch_github_issue_context(issue_id)
+    target = extract_intent_target(context, client)
+
+  bug_exists = verify_bug_reproduction(target, parser)
+  target.is_reproducible = bug_exists
+  if not bug_exists:
+    print(
+        f"[Intent Agent] No bug found: model already parses '{target.expected_split}' correctly."
+    )
+
+  # Step 2: Generator Oversampling
+  assert client is not None
+  raw_candidates = generate_oversample_candidates(
+      target, client, num_candidates=num_candidates, lang=lang)
+  print(
+      f"[Generator Agent] Synthesized {len(raw_candidates)} candidate variations for lang '{lang}'."
+  )
+
+  # Step 3: Base Parser Overlay & Linguistic Expert Pruning
+  aligned_candidates = align_to_base_parser_splits(raw_candidates, target,
+                                                   parser)
+  final_candidates = prune_linguistic_anomalies(
+      aligned_candidates, client, max_keep=max_keep, lang=lang)
+  print(
+      f"[Expert Agent] Pruned to {len(final_candidates)} high-confidence rows.")
+
+  output_lines = final_candidates
+  if outfile:
+    with open(outfile, "w", encoding="utf-8") as f:
+      f.write("\n".join(output_lines) + "\n")
+    print(f"[Staging] Saved {len(output_lines)} lines directly to {outfile}.")
+
+  return output_lines
+
+
+def build_parser() -> argparse.ArgumentParser:
+  """Constructs command-line flags for agentic candidate sample synthesis."""
+  p = argparse.ArgumentParser(description="Agentic candidate sample synthesis.")
+  group = p.add_mutually_exclusive_group(required=True)
+  group.add_argument(
+      "--input", "-i", type=str, help="Target string (e.g. 'いよいよ/はじまる')")
+  group.add_argument(
+      "--issue", type=str, help="GitHub bug report number or URL ID")
+  p.add_argument(
+      "--lang",
+      type=str,
+      default="ja",
+      help="Target model language (default: ja)")
+  p.add_argument(
+      "--num-samples",
+      "-n",
+      type=int,
+      default=30,
+      help="Initial generation count")
+  p.add_argument(
+      "--max-keep", type=int, default=15, help="Pruned final row target")
+  p.add_argument(
+      "--output",
+      "-o",
+      type=str,
+      default="staging_raw.txt",
+      help="Staging destination file")
+  return p
+
+
+def main() -> None:
+  args = build_parser().parse_args()
+  run_agentic_synthesis_pipeline(
+      input_str=args.input,
+      issue_id=args.issue,
+      num_candidates=args.num_samples,
+      max_keep=args.max_keep,
+      lang=args.lang,
+      outfile=args.output,
+  )
+
+
+if __name__ == "__main__":
+  main()
diff --git a/scripts/tests/test_synthesize_samples.py b/scripts/tests/test_synthesize_samples.py
new file mode 100644
index 00000000..e814d556
--- /dev/null
+++ b/scripts/tests/test_synthesize_samples.py
@@ -0,0 +1,168 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Test suite for agentic training data synthesis."""
+
+import os
+import sys
+import tempfile
+import unittest
+from typing import Any
+from unittest.mock import MagicMock, patch
+
+# Module hack to allow importing scripts and budoux from workspace root
+LIB_PATH = os.path.join(os.path.dirname(__file__), '..', '..')
+sys.path.insert(0, os.path.abspath(LIB_PATH))
+
+import budoux  # noqa: E402
+from scripts import synthesize_samples  # noqa: E402
+
+SEP = budoux.utils.SEP  # Canonical character separator '▁'
+
+
+class TestAlignToBaseParserSplits(unittest.TestCase):
+
+  def test_align_to_base_parser_splits_negative_phrase(self) -> None:
+    target = synthesize_samples.IntentTarget(
+        target_phrase="もはや",
+        expected_split="もはや",
+        is_positive=False,
+        is_reproducible=True)
+    mock_parser = MagicMock(spec=budoux.Parser)
+    mock_parser.parse.return_value = ["彼の", "才能は", "もは", "や", "誰にも", "超えられない。"]
+
+    cands = [f"彼の{SEP}才能は{SEP}もはや{SEP}誰にも超えられない。"]
+    result = synthesize_samples.align_to_base_parser_splits(
+        cands, target, mock_parser)
+
+    self.assertEqual(len(result), 1)
+    expected_sentence = f"彼の{SEP}才能は{SEP}もはや{SEP}誰にも{SEP}超えられない。"
+    self.assertEqual(result[0], expected_sentence)
+
+  def test_align_to_base_parser_splits_positive_phrase(self) -> None:
+    target = synthesize_samples.IntentTarget(
+        target_phrase="いよいよはじまる",
+        expected_split=f"いよいよ{SEP}はじまる",
+        is_positive=True,
+        is_reproducible=True,
+    )
+    mock_parser = MagicMock(spec=budoux.Parser)
+    mock_parser.parse.return_value = ["いよいよは", "じまる", "決戦だ。"]
+
+    cands = ["いよいよ▁はじまる決戦だ。"]
+    result = synthesize_samples.align_to_base_parser_splits(
+        cands, target, mock_parser)
+
+    self.assertEqual(len(result), 1)
+    expected_sentence = f"いよいよ{SEP}はじまる{SEP}決戦だ。"
+    self.assertEqual(result[0], expected_sentence)
+
+  def test_align_to_base_parser_splits_discards_missing_target(self) -> None:
+    target = synthesize_samples.IntentTarget(
+        target_phrase="もはや",
+        expected_split="もはや",
+        is_positive=False,
+        is_reproducible=True)
+    mock_parser = MagicMock(spec=budoux.Parser)
+    cands = ["関係のない完全に異なる文章です。"]
+    result = synthesize_samples.align_to_base_parser_splits(
+        cands, target, mock_parser)
+    self.assertEqual(len(result), 0)
+
+
+class TestParseDirectInput(unittest.TestCase):
+
+  def test_parse_direct_input_positive(self) -> None:
+    target = synthesize_samples.parse_direct_input("いよいよ/はじまる")
+    self.assertEqual(target.target_phrase, "いよいよはじまる")
+    self.assertEqual(target.expected_split, f"いよいよ{SEP}はじまる")
+    self.assertTrue(target.is_positive)
+    self.assertTrue(target.is_reproducible)
+
+  def test_parse_direct_input_negative(self) -> None:
+    target = synthesize_samples.parse_direct_input("もはや")
+    self.assertEqual(target.target_phrase, "もはや")
+    self.assertEqual(target.expected_split, "もはや")
+    self.assertFalse(target.is_positive)
+    self.assertTrue(target.is_reproducible)
+
+
+class TestVerifyBugReproduction(unittest.TestCase):
+
+  def test_verify_bug_reproduction_detects_bug(self) -> None:
+    target = synthesize_samples.IntentTarget(
+        target_phrase="いよいよはじまる",
+        expected_split=f"いよいよ{SEP}はじまる",
+        is_positive=True,
+    )
+    mock_parser = MagicMock(spec=budoux.Parser)
+    mock_parser.parse.return_value = ["いよいよは", "じまる"]
+
+    self.assertTrue(
+        synthesize_samples.verify_bug_reproduction(target, mock_parser))
+    mock_parser.parse.assert_called_once_with("いよいよはじまる")
+
+  def test_verify_bug_reproduction_no_bug(self) -> None:
+    target = synthesize_samples.IntentTarget(
+        target_phrase="もはや", expected_split="もはや", is_positive=False)
+    mock_parser = MagicMock(spec=budoux.Parser)
+    mock_parser.parse.return_value = ["もはや"]
+    self.assertFalse(
+        synthesize_samples.verify_bug_reproduction(target, mock_parser))
+
+
+class TestRunAgenticSynthesisPipeline(unittest.TestCase):
+
+  def test_run_pipeline_raises_without_env_or_client(self) -> None:
+    mock_parser = MagicMock(spec=budoux.Parser)
+    with patch.dict(os.environ, {"GEMINI_API_KEY": ""}, clear=True):
+      with self.assertRaises(RuntimeError):
+        synthesize_samples.run_agentic_synthesis_pipeline(
+            input_str="いよいよ/はじまる", client=None, parser=mock_parser)
+
+  @patch("scripts.synthesize_samples.generate_oversample_candidates")
+  @patch("scripts.synthesize_samples.prune_linguistic_anomalies")
+  def test_run_agentic_synthesis_pipeline_end_to_end(
+      self, mock_prune: Any, mock_generate: Any) -> None:
+    mock_client = MagicMock()
+    mock_parser = MagicMock(spec=budoux.Parser)
+    mock_parser.parse.return_value = ["いよいよ", "はじまる", "決戦だ。"]
+
+    raw_sentence = f"いよいよ{SEP}はじまる決戦だ。"
+    mock_generate.return_value = [raw_sentence]
+    expected_overlay = f"いよいよ{SEP}はじまる{SEP}決戦だ。"
+    mock_prune.return_value = [expected_overlay]
+
+    outfile = os.path.join(self.test_dir, "staging_ja.txt")
+    lines = synthesize_samples.run_agentic_synthesis_pipeline(
+        input_str="いよいよ/はじまる",
+        num_candidates=5,
+        max_keep=3,
+        lang="ja",
+        outfile=outfile,
+        client=mock_client,
+        parser=mock_parser,
+    )
+    self.assertEqual(lines, [expected_overlay])
+    self.assertTrue(os.path.exists(outfile))
+
+  def setUp(self) -> None:
+    self.temp_dir_obj = tempfile.TemporaryDirectory()
+    self.test_dir = self.temp_dir_obj.name
+
+  def tearDown(self) -> None:
+    self.temp_dir_obj.cleanup()
+
+
+if __name__ == "__main__":
+  unittest.main()