Current AI benchmarks measure energy consumption at the level of individual model calls, a metric that works fine for single-turn tasks but catastrophically breaks down for agentic systems. When a user asks an AI agent to solve a research-level math problem, that single goal may trigger dozens of internal steps: searching for relevant papers, querying knowledge bases, invoking specialized solvers, and routing results between components. Researchers evaluating these systems today have no unified way to measure the actual energy cost per successful goal completion—they only see raw megajoules per forward pass, creating a measurement gap that masks inefficiency. This blind spot matters because agentic systems are becoming the dominant deployment pattern, yet we're flying without instruments to understand their true resource consumption.
BOHM, a new attribution framework, directly addresses another measurement crisis in compound AI systems. Previous methods, dominated by Shapley-based approaches like SHAP, decompose component contributions by evaluating a coalition value function—a computationally expensive process that requires testing thousands of component combinations to understand which parts of a system actually drive outputs. For hierarchical agent architectures with many specialized components, this overhead becomes prohibitive. BOHM achieves zero-cost hierarchical attribution by eliminating these exhaustive coalition evaluations, instead using gradient information to directly compute marginal contributions. The practical impact is substantial: agents can now introspect on their own decision-making without adding hours of computation, enabling real-time debugging and optimization of multi-stage reasoning pipelines.
A complementary efficiency breakthrough comes from Latent Cache Flow, which eliminates a silent performance drain in model-to-model communication. Current agentic systems pass information between specialized models via text: one model autoregressively decodes its internal state token-by-token, sending plain language to the next model, which then re-encodes it back into latent representations. This round-trip incurs severe latency and loses information in translation. Latent Cache Flow enables direct latent-space communication, allowing models to pass their internal computational state directly without tokenization overhead. In a typical research agent querying multiple specialized solvers, this removes milliseconds of latency per hand-off while preserving full information fidelity—gains that compound across dozens of steps. Together with SciAtlas, a knowledge graph designed to unify fragmented scientific literature for agent use, these papers signal a maturing field learning to measure and optimize the systems it builds.
The Energy per Successful Goal framework crystallizes this shift in thinking. Rather than reporting that training a research math agent costs 500 megajoules, the framework measures energy per solved problem—enabling direct comparison across approaches. An agent that solves ten problems using 600 megajoules proves more efficient than one consuming 400 megajoules to solve three. This goal-level accounting surfaces tradeoffs between exploration, verification, and search that per-invocation metrics hide. As agentic AI moves into production environments where every joule carries cost and carbon weight, these frameworks transform evaluation from noise into signal.