>Isn't a single agent with a good harness better than any multi agent env?

Multiple agents can each have their own isolated LLM context that is focused on their respective narrow scoped tasks. That way, the accumulated "memories", "instructions", etc do not stomp on each other. E.g. the agent loop that is gathering "signals" from the market has a context memory full of instructions related to scraping websites and using REST APIs and it doesn't interfere with the other agent loop that does sentiment analysis that has context memory dedicated to text of news transcripts correlated with historical price signals.

You can have a single agent acting as a top-level "orchestrator" to manage other agents. However, you can't have a single agent that does "everything" for long complicated workflows with high quality. The drawback is the more agents you spawn to take advantage of isolating contexts to improve task completion quality, the more you multiply the tokens consumption and costs.

Another other driver for multi-agents is scaling out cpu usage, especially across different machine boundaries.