M
Marco_Casalaina
In my last post I explored the virtues of a multi-agent system. The "debating agent" pattern (or, as some say, the "maker-checker" pattern) is a fascinating solution to the errors that can arise from a single LLM call. Multi-agents can also use multiple models, allowing for specialized or lower-cost models for some tasks.
With multiple agents, though, comes a fair bit more complexity. There are numerous ways that multiple agents can interact. My Questionnaire Multiagent has a simple, sequential agent pattern - the Question Answerer always goes first, then the Answer Checker, then the Link Checker and then the Manager.
Even with this simple interaction model, though, I've had to work through numerous eventualities; for example, I've had to handle the case where they have an infinitely long argument. And I've had to laboriously prompt engineer all four of them. As the number of agents in your multi-agent system grows, the burden of configuring and maintaining them grows proportionally.
I've recently been experimenting with the use of assistants for both structured and unstructured data, and I've found them to be remarkably powerful on their own. They act sequentially, which makes them easier to debug. They have built-in tools like Code Interpreter, which both writes and executes code, and File Search, which grounds the agent to data. See an example in my 90-second video:
More tools are coming for them, including the ability to call Azure Logic Apps, browse the web, and automatically query various sources of structured data like Microsoft Fabric.
They're comparatively easy to set up, and they're production-tested among our customers and in OpenAI's ChatGPT. They certainly still require data grounding, prompt engineering, bulk evaluation, and care and feeding - but you really can get a lot out of a single agent setup.
So, before you try many AI agents to solve your problem, see if a single agent can get the job done.
For more about AI agents,
Continue reading...
With multiple agents, though, comes a fair bit more complexity. There are numerous ways that multiple agents can interact. My Questionnaire Multiagent has a simple, sequential agent pattern - the Question Answerer always goes first, then the Answer Checker, then the Link Checker and then the Manager.
Even with this simple interaction model, though, I've had to work through numerous eventualities; for example, I've had to handle the case where they have an infinitely long argument. And I've had to laboriously prompt engineer all four of them. As the number of agents in your multi-agent system grows, the burden of configuring and maintaining them grows proportionally.
Before You Try Many Agents, Try Just One
While multi-agent systems garner a lot of attention, single agent systems continue to grow in power. Assistants, now available in Azure OpenAI Service, are a prime example of this.I've recently been experimenting with the use of assistants for both structured and unstructured data, and I've found them to be remarkably powerful on their own. They act sequentially, which makes them easier to debug. They have built-in tools like Code Interpreter, which both writes and executes code, and File Search, which grounds the agent to data. See an example in my 90-second video:
More tools are coming for them, including the ability to call Azure Logic Apps, browse the web, and automatically query various sources of structured data like Microsoft Fabric.
They're comparatively easy to set up, and they're production-tested among our customers and in OpenAI's ChatGPT. They certainly still require data grounding, prompt engineering, bulk evaluation, and care and feeding - but you really can get a lot out of a single agent setup.
So, before you try many AI agents to solve your problem, see if a single agent can get the job done.
For more about AI agents,
- Begin exploring and building your own agents using Azure AI Studio
- Review the Azure OpenAI Assistants quick-start documentation
- Fetch the GitHub repo Questionnaire Multiagent
Continue reading...