Building a Tweet Critique Agent with LangGraph Reflection Pattern

Most tweet generators produce a single draft. If it is weak, you regenerate manually and hope for a better one.

A reflection agent loop is a better pattern: generate, critique, revise, and stop when quality is good enough.

Why Reflection Beats One-Shot Generation #

A one-shot flow usually fails on consistency:

Hooks are generic
Call-to-action is weak
Structure varies across drafts

With reflection, the system can carry forward targeted feedback between iterations instead of restarting from zero context.

Agent Design #

This implementation uses two explicit roles:

Generator: drafts or revises the tweet.
Critic: returns concrete, actionable feedback.

The critic's output is fed back as a message into the next generation turn.

LangGraph Loop Pattern #

LangGraph makes this clean because you can model the loop as a graph with conditional edges.

from langgraph.graph import StateGraph, START, END
 
def should_stop(state):
    # example heuristic: stop after N turns or if score threshold is reached
    if state["iteration"] >= 4:
        return "end"
    if state.get("quality_score", 0) >= 0.9:
        return "end"
    return "continue"
 
builder = StateGraph(dict)
builder.add_node("generate", generate_tweet)
builder.add_node("critique", critique_tweet)
 
builder.add_edge(START, "generate")
builder.add_edge("generate", "critique")
builder.add_conditional_edges("critique", should_stop, {
    "continue": "generate",
    "end": END,
})
 
graph = builder.compile()

Practical Stopping Heuristics #

Do not leave the loop open-ended. Use one or more of these:

Max iterations (for example 3-5)
Character-budget compliance (for example less than 220)
CTA present or not
Hook strength score threshold

Without explicit stop rules, reflection loops can over-optimize and degrade readability. Add hard limits from day one.

What Improves Most in Practice #

In repeated tests, reflection most often improved:

Hook clarity in first 7-10 words
Stronger action-oriented ending
Better formatting for short-form reading
Less passive language

Optional Next Iterations #

Once the basic loop works, useful extensions are:

Multi-variant generation (question, stat, hot-take)
Thread mode (tweet + follow-up)
Lightweight ranker to select best revision
Media prompt suggestions (thumbnail/GIF guidance)

Open Source Reference #

Code repository:

https://github.com/imniteen/Reflection-Tweet-Critique-Agent

Niteen Badgujar