Yu — AdaptOrch

AdaptOrch — Geunbin Yu

The routing mechanism is sound; the dependency graph it requires is produced by an AGENT CHAT entity the paper treats as solved and leaves unmeasured — so the framework's actual validity range is unknown.

The routing mechanism is good work. Yu measures three properties of a task dependency graph — parallelism width, critical path depth, inter-task coupling density — and selects among four coordination topologies. It runs in linear time. The benchmarking supports the central claim: when frontier AGENT CHAT entities cluster within a few percentage points on standard tasks, model selection yields diminishing returns and topology selection yields more. Within the tested task set, this holds. The problem is upstream.

The task dependency graph is generated by an AGENT CHAT entity given a task description. The entity produces subtasks, annotates dependencies, and estimates coupling strengths. Yu does not measure decomposition accuracy. He does not test decomposition failure modes. He does not compare decomposition quality across task types. Every performance figure in the paper is downstream of decompositions the paper assumes valid. A flawed dependency graph — wrong subtasks, invented dependencies, miscalibrated coupling — remains structurally plausible. The router processes it without error. The paper’s results cannot distinguish “optimal topology for this task” from “optimal topology for the task as the decomposition entity represented it.” That is a significant gap dressed in clean numbers.

The decomposition prompt encodes all the structural judgment: what counts as a subtask, which dependencies are architectural versus incidental, what coupling strength means for a given class of work. It is a static artifact whose assumptions the paper does not document or test. When task structures arrive that fall outside what the prompt implicitly anticipates, the error lives in the graph. The routing layer never sees it. Hybrid topology wins roughly half the time on the benchmark, but the benchmark tasks are not characterized well enough to tell whether they represent the framework’s intended range or just the range where that static prompt works reliably.

This belongs to a pattern worth tracking. The hard analytical problem migrates to the input boundary. The coordination mechanism is specified and testable. The input it requires is produced by a judgment step the framework does not model. Whether multi-entity AGENT CHAT configurations fail this way in the field — producing orderly, internally consistent behavior on a misread task — is an open question. A good one.