We continue to analyze the AI Lab papers: Entropy advises to discard half of the sub-agents. Another provocative article from the experience of real applications. The Applied AI team looked at enterprise applications and saw that a single agent with skills often outperformed both the "one prompt for everything" option and the option with a bunch of sub-agents. And usually it comes out cheaper and faster. For one client, a single loop ran about 2x faster than a multi-agent loop. In addition, a dozen specialist sub-agents at the client's agent broke the unified tone of the brand and added delay. In short, everything is like in business: the architecture of the agent copies the organizational structure :) Each team has its own sub-agent, and the orchestrator is placed on top of the platform. What comes out of it:
- Connection is lost when transferring between agents.
- Hundreds of overlapping sub-agents without standards or owners.
–. "Frankenstein-prompt", in which each team added a suggestion.
- Local assessments on which the management makes decisions.
- delays what to do instead:.
- One cycle, one contact window, limitations on moves and budget.
Gradual exposure: the model sees short descriptions of skills and tools and loads them only when necessary. So they reduced the guidance of the cloud code system by more than 80%.
- Group tools by user intent, not by API. After 30-50 tool descriptions, the choice starts to deteriorate.
- Check access rights within the tools and return an understandable denial to the model, then one cycle serves users with different access Sub-agents are still needed if there is measurable degradation from an inflated context, parallel tasks are needed with a new context, or the agent is confused by conflicting tools and instructions. But you need to divide them by context, and not by type of task - otherwise you end up with a broken phone. How to migrate: first examples (how could we live without them); Even 10 pieces will be enough.