Agent Best Practices

Follow these best practices to build reliable, efficient, and maintainable agents.

Model Selection

Use Appropriate Models

Don't use expensive models for simple tasks:

# Good: Use cheaper models for classification
classifier = Agent(model="llama-3")

# Good: Use powerful models for complex reasoning
analyst = Agent(model="gpt-4")

Implement Fallbacks

Always configure fallback models:

{
  "model": {
    "primary": "gpt-4",
    "fallbacks": ["claude-3-sonnet", "gemini-pro"]
  }
}

Memory Management

Set Appropriate TTLs

Configure time-to-live based on data importance:

{
  "memory": {
    "shortTerm": { "ttl": "1h" },
    "longTerm": { "ttl": "30d" }
  }
}

Implement Memory Cleanup

Regularly clean up old memory:

@agent.schedule("0 0 * * *")  # Daily
async def cleanup_memory():
    await agent.memory.cleanup(older_than="30d")

Error Handling

Implement Retry Logic

Handle transient failures:

@agent.on_error
async def handle_error(error):
    if error.is_retryable:
        return await agent.retry(max_attempts=3)
    else:
        await agent.log_error(error)
        return error.fallback_response

Graceful Degradation

Provide fallback responses:

try:
    response = await agent.generate(prompt)
except Exception as e:
    response = "I'm having trouble right now. Please try again."

Performance

Use Caching

Cache common responses:

@agent.cache(ttl=3600)
async def get_faq_response(question):
    return await agent.generate(question)

Optimize Prompts

Keep prompts concise:

# Bad: Verbose prompt
prompt = """
You are a helpful AI assistant. Your job is to...
[500 words of instructions]
Question: {question}
"""

# Good: Concise prompt
prompt = "Answer: {question}"

Security

Validate Input

Always validate user input:

@agent.on_message
async def handle_message(message):
    if not validate_input(message.content):
        return "Invalid input"
    
    return await agent.process(message)

Sanitize Output

Prevent injection attacks:

response = await agent.generate(prompt)
return sanitize(response)

Monitoring

Set Up Alerts

Monitor critical metrics:

swiftclaw alerts create \
  --agent my-agent \
  --metric error-rate \
  --threshold 5% \
  --action notify

Log Important Events

Track agent behavior:

@agent.on_message
async def handle_message(message):
    agent.log("Received message", {
        "user": message.user_id,
        "length": len(message.content)
    })

Testing

Test Locally

Always test before deploying:

swiftclaw dev --test

Use Version Control

Tag releases:

git tag v1.0.0
swiftclaw deploy --version v1.0.0

Following these best practices will help you build reliable, efficient agents that scale.