Can you measure an AI agent's efficiency in separate parts, like memory, tools, and planning, or do they all move together?
Can we decompose agent efficiency into measurable independent components?
This explores whether an agent's efficiency can be split into separate parts, each with its own way of being measured, and where the corpus says that split holds up or breaks down.
This explores whether an agent's efficiency can be split into separate parts, each measured on its own, and where that split holds up. The corpus says yes, into three parts, but the evidence also shows those parts are pushed by the same underlying pressures. Think of them as separate dials on one machine, not three separate machines.
The proposed split is memory, tool use, and planning. Does agent efficiency really break down into three distinct components? argues these are structurally independent, and each has its own cost profile: tokens, latency, and steps. The practical consequence is that improving one axis doesn't automatically improve the others. A compact memory system can still sit next to a wasteful planner, so you have to look at all three.
But independent doesn't mean unrelated. Do efficiency techniques across agent components reveal shared structural constraints? finds that techniques developed separately for each axis keep landing on the same principles: keep the context bounded, make as few external calls as possible, and search in a controlled way instead of open-endedly. That suggests the three axes are views of one constraint: an agent runs in loops, and every loop costs something. Why does agent efficiency differ from model size reduction? makes the same point from the other side. Because agents recurse, making the model cheaper per token matters little next to the total cost of the loop. Efficiency is a success-versus-cost trade-off for the whole system, not a property of one component.
The corpus is most useful on how not to measure it. How should we measure agent system performance beyond task success? shows that identical success rates can hide huge differences in efficiency, reliability, and deployment readiness. It calls for benchmarks that score the harness itself, meaning the trajectory, memory hygiene, and verification cost, not just the final answer. Two other notes cut the problem differently from the three axes. Can multi-agent teams automatically remove their weakest members? scores each agent's contribution within a team and removes the weakest at inference time, which decomposes by team member instead of by function. What predicts success in ultra-long-horizon agent tasks? finds that persistence in feedback loops predicts success better than initial quality, with most models quitting early or burning budget unproductively. That is an efficiency failure the three-axis split doesn't obviously name.
There is also a parallel with reliability. Where does agent reliability actually come from? says reliable agents move memory, skills, and protocols out of the model and into a harness layer. Memory appears in both breakdowns, which hints that the same harness pieces drive both efficiency and reliability. The corpus offers a credible decomposition and a strong case for measuring more than success. It does not demonstrate that the three axes can each be scored cleanly in isolation, and the convergence evidence suggests they can't be fully separated.