diff --git a/secator/ai/actions.py b/secator/ai/actions.py
index 1fd07a099..c2c1def39 100644
--- a/secator/ai/actions.py
+++ b/secator/ai/actions.py
@@ -28,6 +28,8 @@ class ActionContext:
 	"""
 	targets: List[str]
 	model: str
+	api_key: str = ""
+	api_base: str = ""
 	encryptor: Any = None
 	dry_run: bool = False
 	verbose: bool = False
@@ -525,6 +527,50 @@ def _sanitize_child_opts(opts: Any) -> Dict:
 	return clean
 
 
+def build_subagent_prompt(objective: str, targets: list, evidence: str) -> str:
+	"""Wrap the LLM-supplied subagent objective in a structured prompt.
+
+	The `objective` is used verbatim (the parent LLM's intent). `targets` scopes
+	the work; `evidence` (auto-gathered, may be empty) is prior findings the
+	subagent should NOT re-discover.
+	"""
+	targets_str = ", ".join(str(t) for t in targets) if targets else "(inherit parent scope)"
+	evidence_block = evidence.strip() if evidence.strip() else "(none — no prior findings for this scope)"
+	return (
+		f"## Objective\n{objective.strip() or '(no explicit objective given)'}\n\n"
+		f"## Scope\nWork ONLY within these target(s): {targets_str}\n\n"
+		f"## Already known (do not re-run tools that would re-discover these)\n{evidence_block}\n\n"
+		f"## Expected output\nInvestigate the objective, then report your findings concisely. "
+		f"Persist any new findings; do not repeat work already listed under 'Already known'."
+	)
+
+
+def _gather_subagent_evidence(ctx: "ActionContext", targets: list, limit: int = 40) -> str:
+	"""Auto-assemble prior findings for the subagent's targets so it doesn't redo work.
+
+	Queries the workspace (the single source of truth — incl. this run's live findings)
+	for findings whose host/ip/url match any target, capped at `limit`. Best-effort:
+	any failure returns "" (evidence is a nicety, never a blocker).
+	"""
+	targets = [t for t in (targets or []) if t]
+	if not targets:
+		return ""
+	query = {"$or": [{"host": {"$in": targets}}, {"ip": {"$in": targets}}, {"url": {"$in": targets}}]}
+	try:
+		results = ctx.get_query_engine().search(query, limit=limit) or []
+	except Exception:  # noqa: BLE001 - evidence is best-effort; never break the spawn
+		return ""
+	lines = []
+	for r in results[:limit]:
+		d = r.toDict() if hasattr(r, "toDict") else r
+		t = d.get("_type", "finding")
+		key = d.get("url") or d.get("matched_at") or f"{d.get('ip','') or d.get('host','')}"
+		extra = f":{d.get('port')}" if d.get("port") else ""
+		name = f" {d.get('name')}" if d.get("name") else ""
+		lines.append(f"- {t} {key}{extra}{name}".rstrip())
+	return "\n".join(lines)
+
+
 def _run_runner(action: Dict, ctx: ActionContext, runner_type: str) -> Generator:
 	"""Execute a secator task or workflow.
 
@@ -548,6 +594,20 @@ def _run_runner(action: Dict, ctx: ActionContext, runner_type: str) -> Generator
 			return
 		opts["subagent"] = True
 		opts["interactive"] = False
+		# Inherit the parent's resolved LLM config so the subagent can actually run.
+		# Without this it falls back to CONFIG.addons.ai.default_model, which may be a
+		# different provider than the parent (e.g. anthropic-direct vs openrouter) with
+		# no key set -> AuthenticationError before the subagent does anything. setdefault
+		# so an explicit LLM-supplied model/key still wins.
+		opts.setdefault("model", ctx.model)
+		if ctx.api_key:
+			opts.setdefault("api_key", ctx.api_key)
+		if ctx.api_base:
+			opts.setdefault("api_base", ctx.api_base)
+		# 1.b/1.c: structure the subagent's prompt and inject prior findings for its
+		# scope so it doesn't re-run work already done.
+		_objective = opts.get("prompt", "")
+		opts["prompt"] = build_subagent_prompt(_objective, targets, _gather_subagent_evidence(ctx, targets))
 
 	# defense in depth: a spawned runner is never dangerous (CLI --dangerous unaffected)
 	opts["dangerous"] = False
diff --git a/secator/tasks/ai.py b/secator/tasks/ai.py
index 1a0b2c4fd..cdbc87c28 100644
--- a/secator/tasks/ai.py
+++ b/secator/tasks/ai.py
@@ -405,6 +405,8 @@ def _run_loop(self) -> Generator:
 		ctx = ActionContext(
 			targets=self.inputs,
 			model=self.model,
+			api_key=self.api_key,
+			api_base=self.api_base,
 			encryptor=self.encryptor,
 			dry_run=self.dry_run,
 			verbose=self.verbose,
diff --git a/tests/unit/test_ai_actions.py b/tests/unit/test_ai_actions.py
index 84cf0821f..ca5fac58a 100644
--- a/tests/unit/test_ai_actions.py
+++ b/tests/unit/test_ai_actions.py
@@ -557,6 +557,62 @@ def test_run_runner_preserves_existing_session_id(self, mock_build_hooks, mock_t
 		_, kwargs = mock_task_cls.call_args
 		self.assertEqual(kwargs.get('context', {}).get('session_id'), 'from-context')
 
+	@patch('secator.ai.actions.TemplateLoader')
+	@patch('secator.ai.actions.Task')
+	@patch('secator.ai.actions._build_hooks_from_context')
+	def test_run_runner_structures_subagent_prompt(self, mock_build_hooks, mock_task_cls, _tpl):
+		mock_build_hooks.return_value = {'fake': ['hook']}
+		mock_runner = MagicMock(); mock_runner.id = 'r1'; mock_runner.reports_folder = None
+		mock_runner.__iter__.return_value = iter([]); mock_task_cls.return_value = mock_runner
+		ctx = ActionContext(targets=['10.0.0.1'], model='m',
+							context={'workspace_id': 'ws1', 'drivers': ['mongodb']})
+		with patch('secator.ai.actions._gather_subagent_evidence', return_value="- port 10.0.0.1:443"):
+			action = {'action': 'task', 'name': 'ai', 'targets': ['10.0.0.1'],
+					'opts': {'prompt': 'Test auth on the API'}}
+			list(_run_runner(action, ctx, 'task'))
+		_, kwargs = mock_task_cls.call_args
+		prompt = kwargs.get('run_opts', {}).get('prompt', '')
+		self.assertIn('## Objective', prompt)
+		self.assertIn('Test auth on the API', prompt)
+		self.assertIn('- port 10.0.0.1:443', prompt)   # evidence injected
+
+	@patch('secator.ai.actions.TemplateLoader')
+	@patch('secator.ai.actions.Task')
+	@patch('secator.ai.actions._build_hooks_from_context')
+	def test_run_runner_subagent_inherits_parent_llm_config(self, mock_build_hooks, mock_task_cls, _tpl):
+		"""A spawned subagent inherits the parent's resolved model/api_key/api_base so it
+		can actually run (else it falls back to CONFIG.default_model with no key)."""
+		mock_build_hooks.return_value = {'fake': ['hook']}
+		mock_runner = MagicMock(); mock_runner.id = 'r1'; mock_runner.reports_folder = None
+		mock_runner.__iter__.return_value = iter([]); mock_task_cls.return_value = mock_runner
+		ctx = ActionContext(targets=['10.0.0.1'], model='openrouter/anthropic/x',
+							api_key='PARENTKEY', api_base='https://base',
+							context={'workspace_id': 'ws1', 'drivers': ['mongodb']})
+		with patch('secator.ai.actions._gather_subagent_evidence', return_value=""):
+			action = {'action': 'task', 'name': 'ai', 'targets': ['10.0.0.1'], 'opts': {'prompt': 'do x'}}
+			list(_run_runner(action, ctx, 'task'))
+		ro = mock_task_cls.call_args[1].get('run_opts', {})
+		self.assertEqual(ro.get('model'), 'openrouter/anthropic/x')
+		self.assertEqual(ro.get('api_key'), 'PARENTKEY')
+		self.assertEqual(ro.get('api_base'), 'https://base')
+
+	@patch('secator.ai.actions.TemplateLoader')
+	@patch('secator.ai.actions.Task')
+	@patch('secator.ai.actions._build_hooks_from_context')
+	def test_run_runner_subagent_explicit_model_wins(self, mock_build_hooks, mock_task_cls, _tpl):
+		"""An explicit LLM-supplied model on the subagent opts is preserved (setdefault)."""
+		mock_build_hooks.return_value = {'fake': ['hook']}
+		mock_runner = MagicMock(); mock_runner.id = 'r1'; mock_runner.reports_folder = None
+		mock_runner.__iter__.return_value = iter([]); mock_task_cls.return_value = mock_runner
+		ctx = ActionContext(targets=['t'], model='parent/model',
+							context={'workspace_id': 'ws1', 'drivers': ['mongodb']})
+		with patch('secator.ai.actions._gather_subagent_evidence', return_value=""):
+			action = {'action': 'task', 'name': 'ai', 'targets': ['t'],
+					  'opts': {'prompt': 'x', 'model': 'explicit/model'}}
+			list(_run_runner(action, ctx, 'task'))
+		ro = mock_task_cls.call_args[1].get('run_opts', {})
+		self.assertEqual(ro.get('model'), 'explicit/model')
+
 
 @unittest.skipUnless(ADDONS_ENABLED['ai'], 'ai addon not installed')
 class TestSanitizeChildOpts(unittest.TestCase):
@@ -1372,5 +1428,59 @@ def test_child_context_strips_parent_identity_keeps_session(self):
 		self.assertNotIn('scan_id', child)
 
 
+@unittest.skipUnless(ADDONS_ENABLED['ai'], 'ai addon not installed')
+class TestBuildSubagentPrompt(unittest.TestCase):
+	def test_structure_sections_and_objective(self):
+		from secator.ai.actions import build_subagent_prompt
+		p = build_subagent_prompt("Test auth on the API", ["10.0.0.1", "app.x.com"], "- Port 443 open")
+		self.assertIn("## Objective", p)
+		self.assertIn("Test auth on the API", p)          # objective verbatim
+		self.assertIn("## Scope", p)
+		self.assertIn("10.0.0.1", p)
+		self.assertIn("app.x.com", p)
+		self.assertIn("## Already known", p)
+		self.assertIn("- Port 443 open", p)               # evidence injected
+		self.assertIn("## Expected output", p)
+
+	def test_empty_evidence_renders_none(self):
+		from secator.ai.actions import build_subagent_prompt
+		p = build_subagent_prompt("Do X", ["t.com"], "")
+		self.assertIn("(none", p.lower())                 # explicit "none" marker
+
+
+@unittest.skipUnless(ADDONS_ENABLED['ai'], 'ai addon not installed')
+class TestGatherSubagentEvidence(unittest.TestCase):
+	def test_queries_targets_and_formats(self):
+		from secator.ai.actions import _gather_subagent_evidence, ActionContext
+		mock_engine = MagicMock()
+		mock_engine.search.return_value = [
+			{"_type": "port", "ip": "10.0.0.1", "port": 443},
+			{"_type": "url", "url": "http://app.x.com/login"},
+		]
+		ctx = ActionContext(targets=[], model='m', context={'workspace_id': 'ws1'})
+		with patch.object(ctx, 'get_query_engine', return_value=mock_engine):
+			out = _gather_subagent_evidence(ctx, ["10.0.0.1", "app.x.com"], limit=40)
+		# queried by an $or over the targets
+		q = mock_engine.search.call_args[0][0]
+		self.assertIn("$or", q)
+		# formatted a compact summary
+		self.assertIn("port", out)
+		self.assertIn("10.0.0.1", out)
+		self.assertIn("url", out)
+
+	def test_no_targets_returns_empty(self):
+		from secator.ai.actions import _gather_subagent_evidence, ActionContext
+		ctx = ActionContext(targets=[], model='m', context={})
+		self.assertEqual(_gather_subagent_evidence(ctx, [], limit=40), "")
+
+	def test_search_error_returns_empty(self):
+		from secator.ai.actions import _gather_subagent_evidence, ActionContext
+		mock_engine = MagicMock()
+		mock_engine.search.side_effect = Exception("boom")
+		ctx = ActionContext(targets=[], model='m', context={'workspace_id': 'ws1'})
+		with patch.object(ctx, 'get_query_engine', return_value=mock_engine):
+			self.assertEqual(_gather_subagent_evidence(ctx, ["t"], limit=40), "")
+
+
 if __name__ == '__main__':
 	unittest.main()