I set up BabyAGI to scrape mentions and generate weekly summaries for our product. It works, but the pipeline is brittle.
Key components:
- Main agent with a clear objective: "Find all Twitter/X and Reddit mentions of [ProductX] from the last 7 days. Summarize sentiment and key topics."
- Used `serpapi` tool for search, `requests` for direct API calls where possible.
- A second agent to draft the summary post.
The main challenge was filtering noise. Had to refine the task list and add validation steps.
Example task list it created:
1. Search Twitter for recent mentions.
2. Search Reddit subreddits.
3. Filter out irrelevant posts (e.g., same name, different product).
4. Compile list with links.
5. Analyze sentiment per mention.
6. Write summary draft.
Here's the core agent setup:
```python
from babyagi import BabyAGI
from langchain.tools import Tool
# Define tools for search and analysis
tools = [serpapi_tool, sentiment_analysis_tool]
agent = BabyAGI.from_llm(llm=llm, tools=tools, verbose=True)
agent.run(objective=objective, initial_task=initial_task)
```
Biggest pitfall: API costs and rate limiting. You need robust error handling in the task execution loop. Also, the summary quality depends heavily on the LLM you use for the final step.
Results were useful for spotting trends, but not real-time. For a firehose of mentions, you'd need a different architecture.
—cp
—cp