diff --git a/src/control_flow/__init__.py b/src/control_flow/__init__.py
index 4a8e2bcb..7ab48972 100644
--- a/src/control_flow/__init__.py
+++ b/src/control_flow/__init__.py
@@ -1,9 +1,11 @@
 from .settings import settings
 
 # from .agent_old import ai_task, Agent, run_ai
-from .core.flow import Flow
+from .core.flow import Flow, reset_global_flow as _reset_global_flow
 from .core.agent import Agent
 from .core.task import Task
 from .core.controller.controller import Controller
 from .instructions import instructions
 from .dx import ai_flow, run_ai, ai_task
+
+_reset_global_flow()
diff --git a/src/control_flow/core/agent.py b/src/control_flow/core/agent.py
index c43a6464..201eb224 100644
--- a/src/control_flow/core/agent.py
+++ b/src/control_flow/core/agent.py
@@ -17,7 +17,7 @@
 
 
 class Agent(Assistant, ControlFlowModel, ExposeSyncMethodsMixin):
-    name: str = "Agent"
+    name: str
     user_access: bool = Field(
         False,
         description="If True, the agent is given tools for interacting with a human user.",
@@ -42,6 +42,3 @@ async def run_async(self, tasks: list[Task] | Task | None = None):
 
     def __hash__(self):
         return id(self)
-
-
-DEFAULT_AGENT = Agent(name="Marvin")
diff --git a/src/control_flow/core/controller/instruction_template.py b/src/control_flow/core/controller/instruction_template.py
index 3d13d2a3..bedc6ab7 100644
--- a/src/control_flow/core/controller/instruction_template.py
+++ b/src/control_flow/core/controller/instruction_template.py
@@ -65,21 +65,18 @@ class TasksTemplate(Template):
         
         You have been assigned to complete certain tasks. Each task has an
         objective and criteria for success. Your job is to perform any required
-        actions and then mark each assigned task as successful. If a task also
+        actions and then mark each assigned task as successful. If a task
         requires a result, you must provide it. Only work on tasks that are
-        assigned to you. If the task requires a result, do not also post the
-        result in a message, as this would be redundant. Messages should be used
-        only to provide context that is not captured in task results.
-        
-        A "parent" is a task that spawned another task as a subtask. Generally,
-        the subtasks will need to be completed BEFORE the parent task. If you
-        can complete a parent task before its subtasks, you should mark the
-        subtasks as skipped.
-        
-        Tasks have a "depends_on" list of upstream tasks that must be completed
-        before the task itself can be completed. The `mark_success` tool will
-        not be available until all dependencies are met.
+        assigned to you.
         
+        A task may have a "parent", meaning it is a subtask and should be
+        completed before its parent. You can `skip` a subtask if you are able to
+        complete the parent task first.
+
+        A task may have tasks it "depends_on", meaning it must be completed
+        after the tasks it depends on. Tasks will error if you try to complete
+        them before their dependencies are met.
+
         ### Current tasks
         
         At the start of your turn, these are all the tasks being worked on. This
@@ -98,22 +95,23 @@ class TasksTemplate(Template):
         result. For example, a task that asks you to choose, discuss, or perform
         an action must be completed by posting messages before the task is
         marked complete. The objective may require participation from multiple
-        agents. Do not mark a task as complete until the objective is fully met.                
+        agents. Do not mark a task as complete until the objective is fully met.
+        
+        If a task does take a result, do not also write the result in a message. 
         
-        Some tasks may require collaboration with other agents to be completed; others
-        may take you multiple attempts. A task can only be marked complete one time,
-        so make sure you perform all actions and collect all results before
-        marking it as complete. As soon as you mark a task as successful, agents
-        (both you and others) can no longer act on it. 
+        Some tasks may require collaboration with other agents to be completed;
+        others may take you multiple attempts. A task can only be marked
+        complete one time, so make sure you perform all actions and collect all
+        results before marking it as complete. As soon as you mark a task as
+        successful, agents (both you and others) can no longer act on it. 
         
         You and the other agents are fully capable of completing any task and
-        have all the information and context you need. Tasks should only be marked
-        failed due to technical errors like a broken or erroring tool or unresponsive human.
-        Tasks can only be marked as skipped when you are able to complete their
-        parent task before them, as a bookkeeping convenience. Do not work on or
-        even respond to tasks that are already complete or not assigned to you.
-        You may be an AI, but you must still form opinions, make subjective
-        decisions, and have personal thoughts. 
+        have all the information and context you need. Tasks should only be
+        marked failed due to technical errors like a broken or erroring tool or
+        unresponsive human. Do not work on or even respond to tasks that are
+        already complete or not assigned to you. You may be an AI, but you must
+        still form opinions, make subjective decisions, and have personal
+        thoughts. 
         
         """
     tasks: list[Task]
diff --git a/src/control_flow/core/flow.py b/src/control_flow/core/flow.py
index 40b67b4f..a3737f0d 100644
--- a/src/control_flow/core/flow.py
+++ b/src/control_flow/core/flow.py
@@ -1,5 +1,5 @@
 from contextlib import contextmanager
-from typing import Callable, Literal
+from typing import TYPE_CHECKING, Callable, Literal
 
 from marvin.beta.assistants import Thread
 from openai.types.beta.threads import Message
@@ -10,14 +10,32 @@
 from control_flow.utilities.logging import get_logger
 from control_flow.utilities.types import AssistantTool, ControlFlowModel
 
+if TYPE_CHECKING:
+    from control_flow.core.agent import Agent
 logger = get_logger(__name__)
 
 
+def default_agent():
+    from control_flow.core.agent import Agent
+
+    return [
+        Agent(
+            name="Marvin",
+            description="I am Marvin, the default agent for Control Flow.",
+        )
+    ]
+
+
 class Flow(ControlFlowModel):
     thread: Thread = Field(None, validate_default=True)
     tools: list[AssistantTool | Callable] = Field(
         [], description="Tools that will be available to every agent in the flow"
     )
+    agents: list["Agent"] = Field(
+        default_factory=default_agent,
+        description="The default agents for the flow. These agents will be used "
+        "for any task that does not specify agents.",
+    )
     model: str | None = None
     context: dict = {}
 
@@ -48,7 +66,7 @@ def __exit__(self, *exc_info):
         return self.__cm.__exit__(*exc_info)
 
 
-GLOBAL_FLOW = Flow()
+GLOBAL_FLOW = None
 
 
 def get_flow() -> Flow:
diff --git a/src/control_flow/core/task.py b/src/control_flow/core/task.py
index b8d9d7de..770cb855 100644
--- a/src/control_flow/core/task.py
+++ b/src/control_flow/core/task.py
@@ -22,6 +22,7 @@
     model_validator,
 )
 
+from control_flow.core.flow import get_flow
 from control_flow.utilities.context import ctx
 from control_flow.utilities.logging import get_logger
 from control_flow.utilities.prefect import wrap_prefect_tool
@@ -86,9 +87,7 @@ def __repr__(self):
     @field_validator("agents", mode="before")
     def _default_agent(cls, v):
         if v is None:
-            from control_flow.core.agent import DEFAULT_AGENT
-
-            return [DEFAULT_AGENT]
+            v = get_flow().agents
         return v
 
     @field_validator("result_type", mode="before")