A coding orchestrator allowed to delegate chose zero workers, six times out of six

A developer built an open-source orchestration layer for OpenAI Codex in which the supervisor model is allowed to decide, per task, whether spawning worker agents is worth it, and the answer came back the same way every time it was asked ^[1][2][9]. That negative result is more useful than a speedup, because it puts the bottleneck in multi-agent coding where the plumbing already suggested it was: coordination cost, not worker capability.

The system, called Sol-Luna, has two roles ^[1]. A GPT-5.6 instance named Sol owns the task, the decomposition, verification and final review; GPT-5.6 Luna instances act as bounded workers when Sol decides delegation is useful ^[3]. There are two separate adaptive decisions. First, whether to delegate at all, with an optimal worker count that is explicitly allowed to be zero ^[4]. Second, if it does delegate, how much reasoning effort each worker gets, from a ladder of Medium, High, XHigh and Max ^[5].

The control layer is the price sheet. According to the author, parallel workers run in isolated Git worktrees, tasks declare their intended file scope and violations are checked after execution, verification is independently rerun rather than trusting a worker's own PASS result, and Sol still reviews the final output ^[6]. Workers also cannot recursively invoke the orchestrator and grow their own worker trees ^[7]. Every one of those items exists because concurrent agents edit overlapping files, drift outside declared scope, misreport their own results, and collide on Git operations ^[8]. That is the overhead a supervisor is implicitly pricing when it declines.

Earlier benchmarks had already shown parallel Luna workers beating sequential Luna delegation, while one Sol working alone was still faster than both ^[10][11]. The author's hypothesis was that a break-even point existed and could be reached by making the task big enough and genuinely decomposable ^[12]. So the fixtures were built to be favourable to parallelism: a four-module independent workload, a six-module independent workload, and a coupled control where decomposition was deliberately inappropriate ^[13]. The six-module fixture ran to roughly 530 lines of specification and 85 deterministic assertions, about fourteen assertions per module, with each module workable on its own ^[14][15].

One boundary matters. The author says the fixture, though much larger than before, still fit comfortably inside a single Sol session, and that he was testing whether clean parallelism alone produced a crossover, not simulating a multi-hour production repository ^[16][17]. Three modes were compared: delegation disabled, delegation available with Sol free to choose, and delegation forced ^[18]. Nineteen of nineteen runs completed ^[19]. Six were free-choice runs, leaving thirteen in the disabled and forced modes ^[20][21]. In all six free-choice runs, Sol declined to delegate: a delegation rate of zero ^[9][22].

The version of the post available here cuts off before reporting the forced-delegation timings at scale ^[23], so the strong claim is not yet on the table. What the experiment does support is narrower and still worth acting on: inside the range where one model can hold the whole task, the coordination tax exceeds the parallelism gain even when the work is engineered to be independent.

Watch two things. Whether a crossover shows up once the task exceeds a single supervisor session, which is the case this benchmark deliberately did not test ^[16]. And whether the effort ladder ever gets exercised at all, given that it only applies when Sol delegates ^[5], and so far it has not.