Response-Budget Limits in Long-Horizon LLM Agents: A Controlled Study of Task Decomposition

Abstract. Large language model (LLM) agents are increasingly deployed as orchestrated systems — pipelines, parallel branches, planner–executor–reviewer teams — yet the mechanism behind the reported gains is rarely isolated from the models that produce them. This paper tests one specific hypothesis: that a substantial part of the benefit attributed to orchestration is a response-budget effect, in which a single model response can carry only a bounded amount of work, and decomposition succeeds not because it reasons better but because it distributes the same work across more responses. We run a controlled study with a single fixed model (deepseek-flash, hidden reasoning disabled so that the output budget bounds emitted code and nothing else). On a self-built benchmark of 25 utility functions with deterministic unit tests, we manipulate k, the number of functions one response must produce, while holding the per-response budget constant. Per-response success falls monotonically in k, and end-to-end task success collapses from 100% at k=1 to 0% at k=25 — at every budget tested (384, 512, 768 tokens). Truncation, not incorrect code, is the proximate cause: at k ≥ 10 every response exhausts its budget. A second experiment compares five orchestration configurations on the same 25-function task. The single agent scores 0%, and — crucially — so does a loop-plus-recovery agent given three attempts. Pipeline, parallel, and planner–executor–reviewer configurations all reach 100%. The decisive control is that loop-plus-recovery and parallel decomposition consume nearly identical token budgets (5,724 vs. 5,685) for opposite outcomes (0% vs. 100%), isolating orchestration structure rather than compute as the cause. Parallel decomposition is Pareto-optimal among the successful configurations (6.4× lower latency than serial pipeline at equal cost), while the planner–executor–reviewer team spends 2.2× the tokens for no success gain. A budget sweep with hidden reasoning enabled shows that the constraint is not an artefact of a tight budget: one-function responses then consume 5–11× more tokens for essentially the same emitted code and never reach 100% even at 8,192 tokens per response, while twenty-five-function responses truncate at every budget tested. We also report the underlying methodological finding: with reasoning enabled, completion length is dominated by an unpredictable internal chain of thought, and max_tokens ceases to be a valid proxy for the amount of work a response can carry. Artifact. Code and raw JSON results are included as supplementary material. Every figure is rendered directly from the recorded result files; no number in the paper is hand-transcribed. Success is measured by deterministic unit tests, not by an LLM judge.