Do single agents always hit organizational limits?

The survey's central claim is that the limit on agents is organizational, not a shortfall of model power. The abstract lists the requirements of real tasks: "heterogeneous expertise, interdependent subtasks, parallel execution, independent verification, and persistent state." It says these "exceed the organizational capacity of any single agent," and that "simply augmenting an individual agent's capabilities or context cannot resolve this architectural mismatch." The conclusion repeats the point: "a single agent loop struggles to support parallel work, specialized expertise, independent verification, and persistent state."

The paper frames this as the latest step in a sequence of engineering paradigms. Prompt Engineering elicits model capability, Context Engineering manages information access, Harness Engineering organizes external tools and resources, and Loop Engineering enables continual reflection and self-improvement. Each of these improves one agent. The paper's answer is to distribute intelligence across multiple specialized agents and organize it at the system level, which it calls System Intelligence: "the ability of an agent system to organize and coordinate multiple intelligent components into a coherent, adaptive whole that pursues a shared objective." Graph Engineering is the proposed paradigm for doing this, described as one that "uses graph abstractions to make system relations explicit, operational, and adaptable." The literature is sorted into three graph views: work organization, agent coordination, and runtime state management.

This widens the graph framing already in the vault. Can we automatically optimize both prompts and agent coordination? treats nodes as operations and edges as information flow, and shows that one formalism covers prompting techniques and swarms. The survey takes the same structural instinct but organizes it by what the graph represents (tasks, agents, state) rather than by node and edge optimization. Its list of single-agent failures also lines up with Can small language models handle most agent tasks?, which reaches heterogeneous composition from a cost argument rather than a coordination argument. Where the survey treats Context Engineering as one rung on the ladder, Why can language models understand context better than generate it? treats it as a discipline in its own right. The survey's claim is that this discipline, however mature, does not reach the coordination problem.

The excerpt does not show that graph abstractions beat other ways of organizing multi-agent systems, and it gives no example, benchmark, or result for any of the three views. It does not say which of the five task requirements each view addresses. It does not count the costs of multiplying agents, and When does adding more agents actually help systems? is a reminder that coordination adds failure modes of its own. The introduction excerpt only describes the earlier shift toward inference-time engineering. What the passages support is a vocabulary: a way to place existing multi-agent work by what its structure makes explicit, and a claim that single-agent improvements alone will not close the gap.