I describe Level 3 as “The Rise of Collaborative Multi‑Agent Systems”, where a central “Project Manager” agent coordinates specialist agents such as Market Research, Product Design, and Marketing to execute product‑launch workflows.
- Key point: collaborative multi‑agent systems rely on division of labor and coordinated communication.
| Level | Hypothesis |
|---|---|
| 1 | The Emergence of the Generalist Agent |
| 2 | Deep Personalization and Proactive Goal Discovery |
| 3 | Embodiment and Physical World Interaction |
| 4 | The Agent‑Driven Economy |
| 5 | The Goal‑Driven, Metamorphic Multi‑Agent System |
Figure 4 visualizes the five hypotheses about the future of agents.
I illustrate the approach with an example task: “Plan my company's offsite retreat for 30 people in Lisbon next quarter.”
- Open Question: How can we substantially improve LLM reasoning to support reliable, long‑term multi‑agent coordination?
- Open Question: What mechanisms will enable genuine, continual learning between agents within a team?
- Open Question: Which architectural approach (large generalist vs. SLM composition) will dominate or how will they be integrated?
- Open Question: What safety and governance frameworks are needed for agents acting as autonomous economic actors?
- Open Question: How will embodied agents handle real‑world uncertainty and physical safety constraints?
- Key point: present bottlenecks include limited LLM reasoning and nascent inter‑agent learning.
- Decision: focus on team‑based automation rather than a monolithic super‑agent.
- Decision: pursue both generalist agents and Small Language Model (SLM) composition as complementary strategies.