Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs
New framework cuts token waste, conflicts, and wall-clock time in multi-agent systems.
Coordinating teams of large language models has been a binary choice: rigid pre-assigned roles or chaotic free-for-alls. Researchers at MIT, Princeton, and other institutions propose a middle ground with LATTE (Language Agent Teams for Task Evolution). The framework draws on distributed systems principles, having agents collaboratively build and maintain a shared coordination graph that encodes sub-task dependencies, agent assignments, and progress states. This protocol allows agents to dynamically reallocate work, adapt to new tasks, and avoid common pitfalls like contradictory edits, redundant outputs, and communication loops — all without a central controller.
In experiments across multiple collaborative benchmarks and base models (including GPT-4 and open-weight alternatives), LATTE consistently reduced token consumption, wall-clock time, and communication overhead compared to MetaGPT, top-down leader-worker structures, and fully decentralized teams. It also lowered coordination failures — such as file conflicts and repeated outputs — while matching or exceeding task accuracy. For tech professionals deploying multi-agent pipelines, LATTE offers a practical off-ramp from the efficiency-vs-flexibility tradeoff, promising significant cost savings and faster iteration in complex AI workflows.
- LATTE agents dynamically build a shared task graph to track dependencies, assignments, and progress, eliminating the need for fixed roles or pipelines.
- The framework reduces token usage, wall-clock time, and coordination failures (e.g., file conflicts, redundant outputs) versus MetaGPT, Leader-Worker, and static decompositions.
- LATTE matches or exceeds accuracy across diverse base models and collaborative tasks while enabling agents to adapt and discover new work autonomously.
Why It Matters
Slashing token costs and coordination overhead makes multi-agent AI systems more practical and scalable for enterprise workflows.