Human-in-the-loop¶

Human-in-the-loop (or "on-the-loop") enhances agent capabilities through several common user interaction patterns.

Common interaction patterns include:

(1) Approval - We can interrupt our agent, surface the current state to a user, and allow the user to accept an action.

(2) Editing - We can interrupt our agent, surface the current state to a user, and allow the user to edit the agent state.

(3) Input - We can explicitly create a graph node to collect human input and pass that input directly to the agent state.

Use-cases for these interaction patterns include:

(1) Reviewing tool calls - We can interrupt an agent to review and edit the results of tool calls.

(2) Time Travel - We can manually re-play and / or fork past actions of an agent.

Persistence¶

All of these interaction patterns are enabled by LangGraph's built-in persistence layer, which will write a checkpoint of the graph state at each step. Persistence allows the graph to stop so that a human can review and / or edit the current state of the graph and then resume with the human's input.

Breakpoints¶

Adding a breakpoint a specific location in the graph flow is one way to enable human-in-the-loop. In this case, the developer knows where in the workflow human input is needed and simply places a breakpoint prior to or following that particular graph node.

Dynamic Breakpoints¶

Alternatively, the developer can define some condition that must be met for the breakpoint to be triggered. This concept of dynamic breakpoints is useful when the developer wants to halt the graph under a particular condition. This uses a NodeInterrupt, which is a special type of exception that can be raised from within a node based upon some condition.

def my_node(state: State) -> State:
    if len(state['input']) > 5:
        raise NodeInterrupt(f"Received input that is longer than 5 characters: {state['input']}")
    return state

See our guide for a detailed how-to on doing this!

Interaction Patterns¶

Approval¶

Sometimes we want to approve certain steps in our agent's execution.

We can interrupt our agent at a breakpoint prior to the step that we want to approve.

This is generally recommend for sensitive actions (e.g., using external APIs or writing to a database).

With persistence, we can surface the current agent state as well as the next step to a user for review and approval.

If approved, the graph resumes execution from the last saved checkpoint, which is saved to the thread:

# Compile our graph with a checkpoitner and a breakpoint before the step to approve
graph = builder.compile(checkpointer=checkpoitner, interrupt_before=["node_2"])

# Run the graph up to the breakpoint
for event in graph.stream(inputs, thread, stream_mode="values"):
    print(event)

# ... Get human approval ...

# If approved, continue the graph execution from the last saved checkpoint
for event in graph.stream(None, thread, stream_mode="values"):
    print(event)

See our guide for a detailed how-to on doing this!

Editing¶

Sometimes we want to review and edit the agent's state.

As with approval, we can interrupt our agent at a breakpoint prior the the step we want to check.

We can surface the current state to a user and allow the user to edit the agent state.

This can, for example, be used to correct the agent if it made a mistake (e.g., see the section on tool calling below).

We can edit the graph state by forking the current checkpoint, which is saved to the thread.

We can then proceed with the graph from our forked checkpoint as done before.

# Compile our graph with a checkpoitner and a breakpoint before the step to review
graph = builder.compile(checkpointer=checkpoitner, interrupt_before=["node_2"])

# Run the graph up to the breakpoint
for event in graph.stream(inputs, thread, stream_mode="values"):
    print(event)

# Review the state, decide to edit it, and create a forked checkpoint with the new state
graph.update_state(thread, {"state": "new state"})

# Continue the graph execution from the forked checkpoint
for event in graph.stream(None, thread, stream_mode="values"):
    print(event)

See this guide for a detailed how-to on doing this!

Input¶

Sometimes we want to explicitly get human input at a particular step in the graph.

We can create a graph node designated for this (e.g., human_input in our example diagram).

As with approval and editing, we can interrupt our agent at a breakpoint prior to this node.

We can then perform a state update that includes the human input, just as we did with editing state.

But, we add one thing:

We can use as_node=human_input with the state update to specify that the state update should be treated as a node.

The is subtle, but important:

With editing, the user makes a decision about whether or not to edit the graph state.

With input, we explicitly define a node in our graph for collecting human input!

The the state update with the human input then runs as this node.

# Compile our graph with a checkpoitner and a breakpoint before the step to to collect human input
graph = builder.compile(checkpointer=checkpoitner, interrupt_before=["human_input"])

# Run the graph up to the breakpoint
for event in graph.stream(inputs, thread, stream_mode="values"):
    print(event)

# Update the state with the user input as if it was the human_input node
graph.update_state(thread, {"user_input": user_input}, as_node="human_input")

# Continue the graph execution from the checkpoint created by the human_input node
for event in graph.stream(None, thread, stream_mode="values"):
    print(event)

See this guide for a detailed how-to on doing this!

Use-cases¶

Reviewing Tool Calls¶

Some user interaction patterns combine the above ideas.

For example, many agents use tool calling to make decisions.

Tool calling presents a challenge because the agent must get two things right:

(1) The name of the tool to call

(2) The arguments to pass to the tool

Even if the tool call is correct, we may also want to apply discretion:

(3) The tool call may be a sensitive operation that we want to approve

With these points in mind, we can combine the above ideas to create a human-in-the-loop review of a tool call.

# Compile our graph with a checkpoitner and a breakpoint before the step to to review the tool call from the LLM 
graph = builder.compile(checkpointer=checkpoitner, interrupt_before=["human_review"])

# Run the graph up to the breakpoint
for event in graph.stream(inputs, thread, stream_mode="values"):
    print(event)

# Review the tool call and update it, if needed, as the human_review node
graph.update_state(thread, {"tool_call": "updated tool call"}, as_node="human_review")

# Otherwise, approve the tool call and proceed with the graph execution with no edits 

# Continue the graph execution from either: 
# (1) the forked checkpoint created by human_review or 
# (2) the checkpoint saved when the tool call was originally made (no edits in human_review)
for event in graph.stream(None, thread, stream_mode="values"):
    print(event)

See this guide for a detailed how-to on doing this!

Time Travel¶

When working with agents, we often want closely examine their decision making process:

(1) Even when they arrive a desired final result, the reasoning that led to that result is often important to examine.

(2) When agents make mistakes, it is often valuable to understand why.

(3) In either of the above cases, it is useful to manually explore alternative decision making paths.

Collectively, we call these debugging concepts time-travel and they are composed of replaying and forking.

Replaying¶

Sometimes we want to simply replay past actions of an agent.

Above, we showed the case of executing an agent from the current state (or checkpoint) of the graph.

We by simply passing in None for the input with a thread.

thread = {"configurable": {"thread_id": "1"}}
for event in graph.stream(None, thread, stream_mode="values"):
    print(event)

Now, we can modify this to replay past actions from a specific checkpoint by passing in the checkpoint ID.

To get a specific checkpoint ID, we can easily get all of the checkpoints in the thread and filter to the one we want.

all_checkpoints = []
for state in app.get_state_history(thread):
    all_checkpoints.append(state)

Each checkpoint has a unique ID, which we can use to replay from a specific checkpoint.

Assume from reviewing the checkpoints that we want to replay from one, xxx.

We just pass in the checkpoint ID when we run the graph.

config = {'configurable': {'thread_id': '1', 'checkpoint_id': 'xxx'}}
for event in graph.stream(None, config, stream_mode="values"):
    print(event)

Importantly, the graph knows which checkpoints have been previously executed.

So, it will re-play any previously executed nodes rather than re-executing them.

See this additional conceptual guide for related context on replaying.

See see this guide for a detailed how-to on doing time-travel!

Forking¶

Sometimes we want to fork past actions of an agent, and explore different paths through the graph.

Editing, as discussed above, is exactly how we do this for the current state of the graph!

But, what if we want to fork past states of the graph?

For example, let's say we want to edit a particular checkpoint, xxx.

We pass this checkpoint_id when we update the state of the graph.

config = {"configurable": {"thread_id": "1", "checkpoint_id": "xxx"}}
graph.update_state(config, {"state": "updated state"}, )

This creates a new forked checkpoint, xxx-fork, which we can then run the graph from.

config = {'configurable': {'thread_id': '1', 'checkpoint_id': 'xxx-fork'}}
for event in graph.stream(None, config, stream_mode="values"):
    print(event)

See this additional conceptual guide for related context on forking.

See see this guide for a detailed how-to on doing time-travel!

Human-in-the-loop¶

Persistence¶

Breakpoints¶

Dynamic Breakpoints¶

Interaction Patterns¶

Approval¶

Editing¶

Input¶

Use-cases¶

Reviewing Tool Calls¶

Time Travel¶

Replaying¶

Forking¶

Comments