Agents, Agents Everywhere

For this exploration, I’ll use a multi-agent meal planning service. A person can bring up a chat window and generate a meal plan for a family of n, get an individualized daily food schedule optimized for nutrition and energy levels, create a shopping list, schedule a grocery pick-up, and incorporate eating out and changed plans. This is a handy example because I’ve already done some development in this space long before LLMs came along, but also because a single-agent workflow could theoretcally handle the process, while at the same time the space could support both parallel and collaborative workflows, so a multi-agent setup.

Level 0: No Orchestration: a hard-coded workflow

An agent has a prompt, a set of tools and some workflow graph that it follows. Given the (lack of reliability) in current agent workflows, this should be the norm for most deployments.

Level 1: Orchestrator Agent With Sub-Agents as Tools

The obvious routing preference is to have a ‘lead agent’.

Agent orchestration / routing is a piece of work around Random thoughts:

• Managing context for agents in the form of state, memory detail, available tools, hooks and callbacks are a form of knapsack problem where we know we want to minimize context such that we we stay within effective bounds
• Routing is a problem that has to be done right.
• Long-running tasks are

Classifier Model Vs Agent

[1]

• Plain rules engines are not effective agent orchestrators, because the underlying business variables are messy, probabalistic and always changing. Plus ML already exists
• Uses a router agent to field incoming requests and determine where they need to go
• Agent has tools

[2]

• Use action schemas to solidify what can and can’t be done

[3]

• Multi-agents ar handy in research where you can’t predict the states in advance. It can’t be done with a single hard-coded workflow
• Research in particular requires going down rabit holes and spending longer than expected on certain tasks

[4]

• Uses an NLP classifier to judge intent and route tasks

[5]

[6]

• Shows the importance of incorporating real-time feedback and understanding when human intervention is necessary

Appendix

1. Resources:

• 1. Multi Agent Architecture for Advertising at Spotify
• 2. How to Engineer Multi-Agent workflows (Github)
• 3. How We Built Our Multi-Agent Research System
• 4. Agent Squad
• 5. Multi-Agent Collaboration via Evolving Orchestration
• 6. Backpressure is All You Need
• 7. Multi-agent stack reddit thread

2. Definitions:

1. Adaptive Thinking:
2. Interleaved Thinking: