DAF: A Retrospective Paired Case Study of Fixed Agent Workflows

Agent workflows combine model inference with scheduling, dependency transfer, and result storage. These infrastructure choices can affect completion time even when the model and task remain nominally unchanged. This paper presents Darshj’s Agent Framework (DAF), a Rust framework for explicit agent execution, and examines a historical experimental integration against native Codex subagent coordination. In eight paired synthetic ledger workflows using the recorded GPT-6 Astra configuration at medium reasoning effort, the retained scorer marked all eight runs in each arm correct; the median paired native-to-DAF wall-time ratio was 3.65. A separate corrected coding experiment contained two pairs, with all four workflows passing the retained evaluation using the same 187 behavioral cases; its median paired ratio was 2.49. These are descriptive findings from four ledger fixtures and one coding task family. Different coordinator behavior, internal model context, and service conditions prevent attributing the differences to the transport protocol or generalizing them to open-ended software engineering. We distinguish the historical benchmark adapter from the current repository and describe the current implementation’s bounded execution and durable-result contracts. The contribution is an inspectable engineering case study with explicit measurement boundaries, failed-harness disclosure, and a reproducibility agenda, rather than a claim of superior reasoning or universally lower cost. This unreviewed preprint includes a curated benchmark evidence package with 20 scored workflow records, original source digests, historical methods, and an offline consistency and arithmetic verifier. The package does not support a full inference rerun or independent rescoring from task inputs. Source: https://github.com/darshjme/daf/tree/123097cc01796bd9794e6095a7536ad9e47a6e09Contact: [email protected]