Capstone: Local Research Assistant Agent

What this is

No new concepts. This is the final checkpoint of the entire course: a single agent combining every major idea from the Advanced tier (and, underneath, everything from Beginner and Intermediate too). If you can read lesson.py and understand why every piece is there, you've completed this course's full arc, from a single .invoke() call in Lesson 1 to a real, multi-capability agent here.

What it does

A command-line research assistant with:

• A calculator tool (Lessons 13-14)
• RAG search over local notes (Lessons 27-29)
• Real memory across turns via a checkpointer (Lessons 23-24)
• Automatic conversation summarization once a session gets long (Lessons 25-26)
• A human approval gate before an irreversible action (Lesson 30)
• Basic tracing of every tool call (Lesson 34)

Where each piece came from

@tool
def calculator(...): ...
@tool
def search_personal_notes(...): ...

Lessons 13-15 (tool definitions, multiple tools) and 27-29 (loading and splitting notes.txt, embedding it into an InMemoryVectorStore, wrapping similarity search as a tool).

summarizer = SummarizationMiddleware(model=model, trigger=("messages", 12), keep=("messages", 4))

Lessons 25-26's concept (keep old information around in compressed form instead of dropping it), now using LangChain's actual built-in middleware for it, instead of the hand-rolled version from Lesson 26. Once the conversation passes 12 messages, older ones get automatically compressed into a summary, keeping the most recent 4 in full.

agent = create_agent(
    model=model,
    tools=[calculator, search_personal_notes, delete_note_section],
    middleware=[summarizer],
    checkpointer=InMemorySaver(),
    interrupt_before=["tools"],
)

Lessons 23-24 (create_agent, InMemorySaver, thread_id memory) and 30 (interrupt_before, pausing before any tool runs).

The selective approval gate: new here, built from familiar pieces

RISKY_TOOLS = {"delete_note_section"}
...
risky_calls = [c for c in last_message.tool_calls if c["name"] in RISKY_TOOLS]
if risky_calls:
    answer = input("  Approve this action? (yes/no): ").strip().lower()
    if answer not in {"yes", "y"}:
        print("  Rejected, action was not performed.\n")
        return
final = agent.invoke(None, {**config, "callbacks": [tracer]})

Lesson 30's interrupt_before=["tools"] pauses before every tool call, unconditionally, there's no built-in way to say "only pause for this specific tool" at that level.

This capstone adds that selectivity in plain Python: after the agent pauses, we check whether the requested tool's name is in RISKY_TOOLS. If it's not (like calculator or search_personal_notes), we resume immediately, no prompt, the user never notices a pause happened. If it is (delete_note_section), we actually ask before resuming. Same underlying mechanism as Lesson 30, now applied selectively instead of to every tool call uniformly.

Basic tracing, kept lightweight

class BasicTracer(BaseCallbackHandler):
    def on_tool_start(self, serialized, input_str, **kwargs) -> None:
        print(f"  [trace] tool call: {serialized.get('name')}({input_str})")

A stripped-down version of Lesson 34's TracingHandler, just enough to see which tool ran with what arguments, on every turn, without the full timing instrumentation. Real applications tune how much tracing detail they need; this capstone shows the minimum useful amount.