Researchers say most AI agent efficiency gains trace to three components: memory, tool learning, and planning.
A survey paper from arXiv maps the emerging field of agent efficiency, cataloguing approaches that reduce latency, token usage, and step counts without sacrificing task performance. The authors find that despite wide variation in implementation, most techniques converge on a handful of principles: compress and manage context to bound memory costs, shape reinforcement learning rewards to discourage unnecessary tool calls, and use controlled search to prune planning paths. The paper frames the core tension as a Pareto frontier — you can hold cost fixed and push effectiveness up, or hold effectiveness fixed and drive cost down, but there is no free lunch.
That framing matters because agent deployment costs are quietly becoming a bottleneck. Running an agent that loops through dozens of tool calls per task is orders of magnitude more expensive than a single-shot prompt, and no amount of model price cuts changes that arithmetic. A survey that maps where the waste actually lives — memory bloat, redundant tool invocations, inefficient search — gives engineers a principled place to start cutting.
The paper does not claim any of these techniques are solved problems; it flags key challenges and open directions. That is the honest version of progress in a field where most efficiency benchmarks are still being standardized and vendors have every incentive to report only the numbers that flatter their products.