AI Agents with LangGraph & Llama 3 | Control the Execution Flow and State of Your Agent Apps

LangGraph is positioned as a way to control both the execution order and the evolving state of agentic applications—down to loops, branching, and persistence—rather than relying on a model-driven “do this then that” flow. That control matters for building reliable agent apps, because the system can keep a structured state (like chat history and habit data), pause for human input if needed, and repeat tool calls until a task is truly complete.

The walkthrough builds a small habit-tracker agent to demonstrate those capabilities. The app uses three nodes arranged in a loop: an “update habits” node that refreshes the current habit list into state, an “agent” node that reads the user’s latest request along with a system prompt containing the current habits, and a “tools” node that executes whichever tool the model selects. A conditional edge named “should continue” checks whether the agent output includes a tool call; if it does, execution routes to the tools node, then returns to “update habits.” If no tool call is needed, the graph ends and returns a final response. This creates a cycle where the agent can iteratively update or verify progress until completion.

On the data side, the habit tracker is backed by persistence using SQLite. The code defines tables for habits and completion status, then seeds initial habits such as “Feed Damas” (weekends) and “Hit the gym” (Monday, Wednesday, Friday), each with associated tasks. The app exposes tool functions for (1) listing habits for a specific date and (2) adding a new habit with a day-of-week schedule and optional tasks. Each tool is wrapped with parameter schemas and docstrings so the model can call them with structured arguments like day, month, year, habit name, repeat frequency, and task lists.

For the language model, the agent uses Llama 3 via the Groq API (temperature set to 0) and injects a system template that includes the current habits pulled from state. When the user asks to “complete the workout for August 14 2024,” the agent selects the correct habit ID (the gym habit rather than the weekend habit), calls the completion tool with arguments like habit_id=2 and date fields, and then returns a confirmation message. A second call verifies the database state reflects the completion.

The example then scales up: the user adds a new habit—“reading a book every weekday” with a task tag “learning”—and the agent successfully creates it in the database with the expected weekday schedule. Finally, the user requests “scheduled habits for 14th of August 2024,” and the app returns the habits that match that date (including the newly added weekday habit). The key takeaway is that LangGraph’s explicit state and graph edges let the agent reliably map user intent to the right persistent records, while the loop-and-conditional routing ensures tool use happens only when needed and repeats until the job is done.