diff --git a/apn/solver.py b/apn/solver.py
index f3e0d1e..d7783a2 100644
--- a/apn/solver.py
+++ b/apn/solver.py
@@ -34,8 +34,6 @@
 
 logger = logging.getLogger(__name__)
 
-MAX_BACKGROUND_SUBAGENTS = 64
-
 # Deliberately reveals nothing about *why* (no verifier output), so the model
 # cannot search for verifier gaps.
 INCORRECT_MESSAGE = (
@@ -129,9 +127,8 @@ def build_agent(
     """Construct the configured agent loop.
 
     Only exposes functionality common to ``deepagent`` and ``react`` so the two
-    behave identically apart from the loop itself. The ``deepagent``-only
-    settings are the subagent roster, restricted to ``general()``, and
-    background dispatch, capped at ``MAX_BACKGROUND_SUBAGENTS``.
+    behave identically apart from the loop itself. The one ``deepagent``-only
+    setting is the subagent roster, restricted to ``general()``.
     """
     constructor: Callable[..., Agent]
     loop_kwargs: dict[str, Any] = {}
@@ -142,7 +139,6 @@ def build_agent(
         # research() and plan() default to read-only sandbox tools (read_file,
         # list_files, grep), and this is not made clear to agents.
         loop_kwargs["subagents"] = [general()]
-        loop_kwargs["background"] = MAX_BACKGROUND_SUBAGENTS
     elif agent_type == "react":
         constructor = react
     else:
diff --git a/configs/bloom68.yaml b/configs/bloom68.yaml
index 80cb3cd..4ef26b0 100644
--- a/configs/bloom68.yaml
+++ b/configs/bloom68.yaml
@@ -27,7 +27,7 @@ models:
 
 # 1 model x 68 samples x $300/sample cap = $20,400 max; gated runs sit near max.
 cost_limit: 300.0
-working_limit: 129_600  # 36 * 60 * 60
+working_limit: 259_200  # 72 * 60 * 60
 
 # $ per 1M tokens; keys must match Inspect's resolved <provider>/<model> names.
 model_cost_config:
diff --git a/tests/test_agent.py b/tests/test_agent.py
index 8e570f8..48ed3b0 100644
--- a/tests/test_agent.py
+++ b/tests/test_agent.py
@@ -106,3 +106,9 @@ def test_deep_agent_offers_only_general_subagent(
     # alone, since it is the only one that inherits the parent's Lean tools.
     seen = _build(monkeypatch, "deep")
     assert [sa.name for sa in seen["subagents"]] == ["general"]
+
+
+def test_react_agent_gets_no_subagents_kwarg(monkeypatch: pytest.MonkeyPatch) -> None:
+    # react() has no subagents parameter; the deep-only kwarg must not leak.
+    seen = _build(monkeypatch, "react")
+    assert "subagents" not in seen