Agents d’autocorrection et de réflexion
Apprenez à construire des agents capables d’évaluer de manière critique leurs propres sorties et d’affiner progressivement leurs réponses ou leurs actions.
Agents d’autocorrection et de réflexion est une leçon AI Agents with LangChain & Autonomous Workflows gratuite sur CoddyKit. Ceci est la leçon 3 sur 6. Tu peux lire la leçon complète ci-dessous gratuitement — puis la pratiquer en direct dans le navigateur avec un éditeur de code intégré et un tuteur IA 24/7. Elle fait partie du parcours d'apprentissage AI Agents with LangChain & Autonomous Workflows, et ta progression se synchronise sur le web et l'application CoddyKit. Le cours AI Agents with LangChain & Autonomous Workflows comprend 6 leçons au total.
Certaines parties de cette leçon n'ont pas encore été traduites et s'affichent en anglais.
Agents That Think Twice
Ever wished your AI agent could check its own work? That's exactly what self-correction and reflection agents do!
These agents are designed to not just generate an output, but also to critically evaluate it and make improvements. Think of it as an agent having an internal "editor" or "critic".
The Reflective Cycle
The core idea of reflection is an iterative process:
- Generate: The agent produces an initial response or action.
- Reflect: It then uses another prompt (or a different LLM) to analyze its own output against specific criteria.
- Refine: Based on the reflection, the agent adjusts its original output, aiming for better quality or accuracy.
This cycle can repeat multiple times until a satisfactory result is achieved.
Inside a Reflection Agent
A self-correcting agent typically involves a few key parts:
- Generator LLM: The primary Large Language Model that creates the initial output.
- Critic/Reflector LLM: Often the same LLM, but prompted to act as a critic, evaluating the generator's output.
- Reflection Prompt: A carefully designed prompt that guides the critic LLM on what to evaluate and how.
- Refinement Mechanism: The logic that takes the critic's feedback and uses it to improve the next generation.
The Generator Prompt
Just like any LLM interaction, a clear initial prompt for the "generator" is crucial. This prompt tells the agent what task to perform.
For example, if we want an agent to write a short story, the generator prompt would define the story's theme, characters, and length. The goal is a solid first draft.
prompt_template = "Write a 3-sentence story about a brave knight and a dragon."Crafting the Reflection Prompt
The reflection prompt is where the magic happens. It instructs the LLM to review the previous output and identify areas for improvement.
This prompt should include specific criteria. For our story example, it might ask: "Is the story engaging? Does it clearly mention a knight and a dragon? Is it exactly 3 sentences?"
reflection_prompt = "Critique the following story: '{story}'. Does it meet these criteria: 1. Engaging? 2. Knight and Dragon present? 3. Exactly 3 sentences? Provide specific feedback for improvement."Simple Reflection Example
Let's trace our story example conceptually:
- Generate: LLM writes an initial story draft.
- Reflect: LLM reads the story and the reflection prompt, then gives feedback (e.g., "It's 4 sentences, not 3.").
- Refine: The agent receives the feedback. It then tries to rewrite the story, incorporating the feedback (e.g., shortening it to 3 sentences).
This loop continues until the story meets all criteria or a limit is reached.
Building a Refinement Loop
In practice, the refinement process often involves a loop. The agent continues to generate and reflect until a set condition is met. This condition could be:
- A maximum number of iterations.
- The reflection feedback indicates no further improvements are needed.
- A specific quality metric is achieved.
This iterative process allows for continuous self-improvement, making agents more robust.
Code: A Simple Reflection Loop
Here's a simplified Python example demonstrating a basic self-correction loop. We'll use a placeholder LLM for clarity, focusing on the generate, reflect, and refine logic.
class MockLLM:
def __init__(self):
pass
def generate_initial_output(self, task_prompt):
# Simulate a generator LLM that sometimes gets it wrong
if "start with 'Hello'" in task_prompt:
return "Greetings! How are you?" # Incorrect initial attempt
return "Hello World!" # Correct initial attempt based on prompt
def reflect_and_suggest(self, original_task, generated_output):
# Simulate a critic LLM
reflection_prompt = (
f"Given the task: '{original_task}'. "
f"The generated output was: '{generated_output}'. "
"Critique this output: Does it start with 'Hello'? "
"If not, suggest a correction. Be concise."
)
# Simplified reflection logic for the mock LLM
if not generated_output.strip().startswith("Hello"):
return "Doesn't start with 'Hello'. Suggestion: Start with 'Hello'."
return "Output meets criteria."
# --- Main Logic ---
llm = MockLLM()
task = "Create a greeting sentence that starts with 'Hello'."
current_output = llm.generate_initial_output(task)
print(f"Initial Output: {current_output}")
for i in range(3): # Max 3 attempts (initial + 2 refinements)
reflection = llm.reflect_and_suggest(task, current_output)
print(f"Reflection {i+1}: {reflection}")
if "Output meets criteria" in reflection:
print("Agent self-corrected successfully!")
break
else:
# Simulate applying the suggestion
if "Start with 'Hello'" in reflection and not current_output.startswith("Hello"):
current_output = "Hello" + current_output[current_output.find(' '):] # Simple fix
print(f"Refined Output: {current_output}")
else:
print("No specific refinement strategy for this feedback. Stopping.")
break
print("\nFinal Output:", current_output)Benefits & Trade-offs
Reflection agents are powerful for tasks requiring high accuracy or complex reasoning, especially when a single pass might miss nuances.
- Improved Accuracy: Reduces errors and hallucinations.
- Robustness: Handles complex prompts better.
- Quality: Leads to higher-quality outputs.
However, they increase computational cost and latency due to multiple LLM calls. Use them where quality justifies the overhead.
Test Your Knowledge
Consider an agent designed to summarize a document and then self-correct. What is the primary purpose of the "reflection prompt" in this scenario?
Recap: Self-Correction
We've explored how self-correction and reflection agents work. These agents enhance AI capabilities by allowing them to critically evaluate and refine their own outputs.
- They follow a generate, reflect, refine cycle.
- Key components include a generator, a critic, and specific reflection prompts.
- While increasing cost and latency, they significantly boost output quality and accuracy for complex tasks.
This ability to "think twice" is a crucial step towards more robust and intelligent AI systems!
Questions Fréquemment Posées
La leçon « Agents d’autocorrection et de réflexion » est-elle gratuite ?
Oui — le texte complet de « Agents d’autocorrection et de réflexion » est gratuit à lire ici sur le web. Pour la pratiquer de manière interactive (un éditeur de code intégré et un tuteur IA 24/7) et déverrouiller le reste du cours AI Agents with LangChain & Autonomous Workflows, passe à CoddyKit PRO. Le cours AI Agents with LangChain & Autonomous Workflows comprend 6 leçons au total.
Qu'est-ce que j'apprendrai dans « Agents d’autocorrection et de réflexion » ?
Apprenez à construire des agents capables d’évaluer de manière critique leurs propres sorties et d’affiner progressivement leurs réponses ou leurs actions. Tu pratiques AI Agents with LangChain & Autonomous Workflows avec du code pratique que tu exécutes directement dans le navigateur, et un tuteur IA 24/7 répond à tes questions au fur et à mesure que tu avances dans la leçon.
Dois-je avoir de l'expérience pour commencer AI Agents with LangChain & Autonomous Workflows ?
Aucune expérience préalable n'est requise. AI Agents with LangChain & Autonomous Workflows sur CoddyKit est structuré pour les débutants jusqu'aux apprenants avancés, donc tu peux commencer ici ou depuis le début et avancer à ton rythme. Ceci est la leçon 3 sur 6.
Combien de temps prend la leçon « Agents d’autocorrection et de réflexion » ?
La plupart des leçons CoddyKit prennent environ 5–10 minutes. Chacune est courte et interactive, tu progresses régulièrement et tu repiques exactement où tu t'es arrêté sur le web et l'app.
Peux-tu écrire et exécuter du code dans cette leçon AI Agents with LangChain & Autonomous Workflows ?
Oui. Chaque leçon AI Agents with LangChain & Autonomous Workflows inclut un éditeur de code intégré, tu écris et exécutes du vrai code directement dans ton navigateur et tu reçois des retours IA instantanés — aucune configuration locale requise.
Toutes les leçons de ce cours
- Agents ReAct et planification-exécution
- Conceptions hiérarchiques des agents
- Agents d’autocorrection et de réflexion
- Architectures cognitives pour les agents
- Modèles de collaboration multi-agents
- Systèmes d’agents hybrides