Research area 03
Multi-agent and agentic systems
Architectures in which several model-driven agents plan, delegate, negotiate, and act, and the coordination, safety, and cost questions that come with them.
- agents
- coordination
- orchestration
- cooperative-ai
Agentic AI connects language models to tools, memory, and other agents so they can carry out multi-step work. The classic multi-agent systems literature, with its focus on coordination, negotiation, and robust autonomy, is suddenly practical again, and the open question is which of its lessons transfer to agents built on large models.
Our work treats multi-agent designs as engineering trade-offs. Decomposition can improve transparency and modularity; it can also multiply cost and create failure modes that no single agent would have. We build and measure.
What we work on
- Workflow design. Planner, implementer, and reviewer roles; adversarial review agents; evidence labels that travel with every claim an agent makes.
- Coordination under constraints. Budgets, timeouts, and escalation rules that keep a system of agents predictable.
- Human-in-the-loop by design. Where the human decides, how the system asks, and how it is overruled.