Agent Teams
Coordination, communication, and cost efficiency in multi-agent systems, so that teams of agents can take on long-horizon work that no single agent finishes reliably.
Many agent workloads are already multi-agent in practice: a task is split into subtasks that run in parallel, each with its own tools, context, and, for computer use, its own virtual machine. The hard part is coordination. Sub-agents must share what they learn without sharing state that cannot be merged, the overall plan has to stay coherent as branches finish at different times, and the whole team has to cost less than the single-agent alternative it replaces.
Spine-Branch (Zhang et al., 2026) is a coordination scheme for multi-agent computer use built around exactly that constraint: a spine agent carries the main task on one continuously evolving machine, while branch agents run in parallel on disposable machines to gather what the spine needs, so no machine state ever has to be merged. On 200 long-horizon tasks and three agent backbones it raises success rates while cutting per-task cost substantially. Efficient multi-agent systems more broadly, including how teams of agents are trained and evaluated, are an active thread of my team’s work; see also Agentic RL Training Recipes and Agent Benchmarks and Evaluation.