Best Practices

Agent Design Patterns

News Scraper + Summarizer

Start (cron: "0 9 * * *")
  -> HTTP API (fetch news API)
  -> Agent (summarize top stories)
  -> End

Social Media Monitor

Start (mention trigger)
  -> Agent (classify intent: question, complaint, praise, spam)
  -> If/Else (is spam?)
    -> True: End
    -> False: Agent (generate response with brand memory)
  -> End

Research Assistant

Start (manual)
  -> Agent (break down query into sub-questions)
  -> Parallel Fork
    -> HTTP API (source 1)
    -> HTTP API (source 2)
  -> Parallel Join
  -> Agent (synthesize findings)
  -> End

Code Review Bot

Start (webhook from GitHub)
  -> HTTP API (fetch PR diff)
  -> Agent (review code for bugs and style)
  -> End

Cost Optimization

1. Use cheaper models for simple tasks — gpt-4o-mini handles most classification and formatting
2. Limit chat history — Set chatHistoryLimit to 10-20 messages instead of unlimited
3. Use routing strategies — Let the system pick the optimal model per request
4. Set budget limits — Configure daily/monthly budgets to prevent runaway costs
5. Cache with brand memory — Pre-load common knowledge instead of re-fetching

Common Mistakes

• Overly broad system prompts — Be specific about what the agent should do
• Missing error handling — Use guardrail nodes to catch bad inputs
• No budget limits — Always set daily limits on production agents
• Too many tool iterations — Set maxToolIterations to prevent infinite loops
• Ignoring chat history costs — Long conversations can be expensive with large models