FlowForge: Dual Systems in AI & CFD

• FlowForge is a dual-use term representing an interactive multi-agent workflow design tool with visual guidance and a staged local rollout engine for CFD flow prediction.
• The workflow variant employs a three-level hierarchical design space with design pattern visualization to streamline multi-agent system creation and improve performance metrics.
• The CFD variant uses a staged autoregressive local predictor with a compiled traversal plan to achieve low RMSE, robust error containment, and efficient execution.

Searching arXiv for the specified FlowForge papers to ground the article. FlowForge is a research name used for two distinct systems in the arXiv literature. In human–AI workflow design, “FlowForge: Guiding the Creation of Multi-agent Workflows with Design Space Visualization as a Thinking Scaffold” denotes an interactive visualization tool for constructing multi-agent workflows through structured design-space exploration and in-situ guidance (Hao et al., 21 Jul 2025). In computational fluid dynamics, “FlowForge: A Staged Local Rollout Engine for Flow-Field Prediction” denotes a surrogate modeling framework that predicts future flow fields through a compiled locality-preserving update schedule executed by a lightweight local predictor (Zhang et al., 21 Apr 2026). The shared name therefore refers not to a single research lineage, but to two unrelated artifacts that each formalize and scaffold a complex design or inference process within their respective domains.

1. Dual usage of the name

The term FlowForge appears in two unrelated 2025–2026 arXiv works with different problem settings, architectures, and evaluation criteria. One addresses the design of multi-agent large-language-model workflows; the other addresses autoregressive prediction of CFD flow fields. The coincidence of naming is notable because each system operationalizes “flow” differently: workflow composition in one case, and physical flow-field rollout in the other (Hao et al., 21 Jul 2025).

Variant	Domain	Core definition
FlowForge	Multi-agent workflow design	An interactive visualization tool for creating multi-agent workflows through structured visual exploration and in-situ guidance
FlowForge	Flow-field prediction	A staged local rollout engine that predicts future flow fields by compiling a locality-preserving update schedule and executing it with a shared lightweight local predictor

This dual usage can create bibliographic ambiguity. A plausible implication is that citations to “FlowForge” require explicit disambiguation by title or arXiv identifier, since the two systems do not share methodology, task formulation, or empirical benchmarks.

2. FlowForge for multi-agent workflow design

In the multi-agent systems literature, FlowForge is designed to help AI practitioners systematically explore and build multi-agent workflows by explicitly visualizing the vast design space of possible solutions and providing in-situ guidance drawn from well-established workflow design patterns (Hao et al., 21 Jul 2025). Its motivation is grounded in four challenges identified in formative interviews with expert workflow designers: navigating a vast, high-dimensional design space; an unstructured, trial-and-error approach that is both time-consuming and prone to fixation on early solutions; balancing competing performance metrics such as latency, computational cost, and accuracy; and a lack of practical, tool-embedded guidance connecting design patterns to concrete implementations.

The system adopts a client-side web application architecture implemented in React for the UI, D3 for custom glyphs and scatterplots, and ReactFlow for canvas node-link editing, while orchestrating workflow generation and execution via LangGraph. Internally, it maintains a hierarchical graph $G=(V,E)$ of candidate workflows $v\in V$, where each edge $(u\to v)\in E$ represents refinement from a parent design $u$ to a more concrete child $v$. Users submit a task description, and the system issues prompt-based queries to GPT-4o to generate initial decompositions at Level 1, design-pattern assignments at Level 2, and agent-level optimizations at Level 3. Workflow metadata, including structural description, estimated metrics, and pattern assignments, are stored in the graph and rendered in two coordinated views: a Design Space View and a Canvas View.

The workflow-design process is organized into three hierarchical levels ranging from abstract to concrete. At Level 1, task planning, users explore alternative decompositions of the input task into subtasks. At Level 2, agent assignment, the system generates suggested agent-assignment patterns such as Reflection and Supervision. At Level 3, agent optimization, the workflow becomes fully executable and users tune prompts, change the LLM model, attach tools such as a web-search API, or connect to retrieval databases. This hierarchy is central to the paper’s claim that users can move from high-level planning to detailed implementation while comparing alternative solutions across multiple performance metrics (Hao et al., 21 Jul 2025).

3. Visualization model and guidance mechanisms in workflow FlowForge

The multi-agent FlowForge formalizes its design space as a directed acyclic graph $G=(V,E)$ in which each vertex $v$ is a complete workflow at some abstraction level $\ell(v)\in\{1,2,3\}$ and carries a performance-metric vector

$$\mathbf{m}(v) = \bigl(\text{cost}(v),\,\text{latency}(v),\,\text{accuracy}(v),\,\dots\bigr)\in\mathbb{R}^d.$$

Edges exist only when $\ell(v)=\ell(u)$ for lateral exploration or $v\in V$0 for refinement. Cost and latency are estimated prior to execution by counting the number of LLM calls and applying known per-call estimates:

$v\in V$1

where $v\in V$2 are vendor-provided billing coefficients and $v\in V$3 is average round-trip time. Users may also supply manual or post-hoc annotations for subjective dimensions such as “creativity score” (Hao et al., 21 Jul 2025).

The Design Space View combines a hierarchical tree and a 2D scatterplot. In Level 1, tree nodes are rendered as glyphs in which the number of arcs indicates the count of subtasks and the radius reflects the length of a sequential chain. In Level 2, glyphs become bar charts over a pseudo-time axis, with each bar’s height encoding the number of concurrent agent invocations in that step. In Level 3, the glyph is replaced by a file-icon indicating that the workflow is fully executable. The scatterplot renders workflows at coordinates $v\in V$4 for user-chosen dimensions, with color and glyph size encoding abstraction level and cost.

The in-situ guidance mechanism embeds eight core design patterns drawn from a literature synthesis of 43 multi-agent workflows. The patterns listed in the paper are, at Level 1, Sequential and Parallel; at Level 2, Reflection, Redundant, Supervision, and Discussion; and at Level 3, Single-Agent augmentations. Pattern recommendation uses a suitability score

$v\in V$5

where each feature $v\in V$6 is estimated by GPT-4o in its recommendation prompt and the weights $v\in V$7 are tuned to reflect user priorities. When a metric such as latency is selected on a scatterplot axis, the axis is annotated with the patterns most strongly correlated with low, medium, and high values of that metric; the paper gives “Redundant at the low-latency end” and “Discussion at the high-latency end” as examples (Hao et al., 21 Jul 2025). This suggests that the tool treats design patterns not only as reusable workflow templates but also as explanatory factors in trade-off analysis.

4. Empirical use and evaluation of workflow FlowForge

The paper reports two usage scenarios. In “Fast-Forward Video Planning,” a researcher seeks to generate a 25-second promotional video for a paper. Five alternative decompositions with 3–5 subtasks appear, spanning conventional and storytelling modes. The selected decomposition parallelizes “explain method” and “highlight impact” to increase creative narrative. At Level 2, plotting “creativity score” on the y-axis shows Discussion and Redundant patterns clustering at higher creativity. The user replaces a single-agent opening with a 3-agent Discussion, sets the number of discussion rounds to 3, fine-tunes prompts at Level 3, executes the workflow, and rates the resulting five candidate scripts; the updated ratings appear automatically on the scatterplot (Hao et al., 21 Jul 2025).

In “Data Storytelling from a Dataset,” the user supplies a movie-dataset JSON and explores decompositions with 2–4 subtasks such as “identify trend,” “choose visualization type,” and “narrate insight.” Simpler decompositions are observed to yield low latency but sometimes irrelevant stories. By plotting “topic coherence” against “runtime,” the user identifies a Pareto front of four workflows. The slower workflow benefits from Reflection, while the faster one uses Parallel; Level 3 tuning of the critic’s scoring prompt further boosts coherence.

The formal user evaluation is a within-subjects observational study with $v\in V$8 comparing FlowForge against LangGraph Studio, described as the state-of-the-art visual IDE without scaffolding. Participants, all experienced in multi-agent LLM workflows but new to both tools, completed a 25-minute task of designing a 25-second paper video workflow. The reported quantitative findings are specific: time to first runnable workflow was $v\in V$9 min for FlowForge versus $(u\to v)\in E$0 min for the baseline, with $(u\to v)\in E$1 and $(u\to v)\in E$2; the number of distinct workflows explored was 3.2 versus 1.2 on average; FlowForge users employed all five Level 2 patterns, whereas the baseline used only single-agent workflows; and post-study Likert ratings favored FlowForge on ease of use and confidence while yielding lower reported need for technical support. Qualitative observations attribute these outcomes to hierarchical guidance, broader divergent exploration at Level 1, reduced premature fixation, and transparent trade-off reasoning via the linked scatterplot (Hao et al., 21 Jul 2025).

5. FlowForge for staged local rollout in flow-field prediction

In computational fluid dynamics, FlowForge is introduced as a staged local rollout engine for future flow-field prediction that rewrites spatial sites stage by stage rather than producing the next frame in a single global pass (Zhang et al., 21 Apr 2026). Its architecture consists of an offline compiler and an online executor. The compiler fixes a rollout plan comprising a traversal order $(u\to v)\in E$3 of the $(u\to v)\in E$4 grid sites, a partition of $(u\to v)\in E$5 into $(u\to v)\in E$6 sequential stages $(u\to v)\in E$7, and for each site a bounded list of at most $(u\to v)\in E$8 earlier sites from which it may read. The executor maintains a working buffer $(u\to v)\in E$9, initialized from the last observed field $u$0, and overwrites it stage by stage. Within each stage, all sites in $u$1 are updated in parallel by the same lightweight local predictor $u$2, conditioned on the static history $u$3 and values in $u$4 written by earlier stages.

The method is motivated by the claim that one-step physical dependencies are local because of finite propagation speed and diffusion radius. In contrast to global models with full-field attention or spectral operators, which are described as having higher latency and being fragile under input defects, FlowForge enforces the spatial autoregressive factorization

$u$5

and amortizes the sequential chain into $u$6 stages (Zhang et al., 21 Apr 2026). The stated objective is to balance parallelism, locality, and error containment.

The paper’s locality analysis specifies a required dependency radius

$u$7

with

$u$8

and states that on a grid with spacing $u$9 it suffices to cover $v$0 rings in $v$1 shells. The plan ensures that for each $v$2, at least $v$3 of its $v$4-ring neighbors lie in earlier stages. The lightweight local predictor is a 4-layer MLP with approximately $v$5 parameters. Its inputs per site are static features $v$6, context features $v$7 from up to $v$8 neighbors in earlier stages, and positional encoding $v$9; the output is the predicted next-step state $G=(V,E)$0 (Zhang et al., 21 Apr 2026).

6. Experimental profile and theoretical claims of rollout FlowForge

The rollout FlowForge is evaluated on CFDBench, PDEBench, and BubbleML. The listed datasets are CFDBench incompressible Navier–Stokes cases at $G=(V,E)$1 resolution—Cavity, Tube, Dam, and Cylinder; PDEBench cases at $G=(V,E)$2—Diff-React (FitzHugh–Nagumo), Rand-M0.1, and Rand-M1.0; and BubbleML mixed-resolution cases up to $G=(V,E)$3—PB-Gravity, PB-Subcooled, FB-Gravity, and FB-VelScale (Zhang et al., 21 Apr 2026). Training uses a single RTX-3090 GPU, Adam with learning rate $G=(V,E)$4, MSE loss, StepLR, up to 150 epochs or a 48h cap, and batch size tuned per method to saturate 24 GB VRAM. Robustness protocols include additive Gaussian noise with $G=(V,E)$5 and spatial masking with random blocks or edge masks of area fraction $G=(V,E)$6. Multi-step rollout predicts up to 100 future frames autoregressively.

The reported quantitative summary states that FlowForge is best or second-best in RMSE on 10 of 11 datasets and routinely has the lowest divergence error on 10 of 11 datasets. On lid-driven cavity, where the paper notes strong global pressure coupling, global models are slightly better, and this is presented as a limit of purely local rollout. The method is further reported to exhibit gently growing rollout RMSE with divergence remaining near zero, whereas U-Net and FNO drift more rapidly in multi-step prediction. Under corruption, the averaged normalized RMSE increase is lowest across nearly every noise and masking regime, and qualitative examples are said to show that global models imprint artifact patterns far from corruptions while FlowForge errors remain locally confined. Latency is described as having per-point inference time that remains flat or slightly decreasing as grid resolution grows, in contrast to mid-scale latency spikes for U-Net and FNO and linear scaling in $G=(V,E)$7 for DeepONet (Zhang et al., 21 Apr 2026).

The ablations emphasize that staged execution rather than locality alone is critical. A local one-shot predictor with the same local context but no staged rollout is reported to have very poor accuracy, with the example “RMSE $G=(V,E)$8 vs $G=(V,E)$9.” Stage count $v$0 mediates a latency–accuracy trade-off: fewer stages yield less context and higher RMSE, while more stages yield richer context and lower RMSE but longer sequential depth; latency remains nearly constant until $v$1 becomes very large. Traversal order, including outward spiral, raster scan, Hilbert curve, and random orderings, gives nearly identical performance once $v$2 and $v$3 ensure coverage. Theoretical support is given in three forms: causality by construction leading to a block lower-triangular Jacobian with no feedback within a stage; an error-propagation bound in which perturbations attenuate by $v$4 under stage-wise Lipschitz constants $v$5; and a coverage bound controlling one-step truncation error through the tail-mass of the advection–diffusion kernel (Zhang et al., 21 Apr 2026).

7. Cross-cutting interpretation and scope

Although they are unrelated systems, both works named FlowForge impose explicit intermediate structure on spaces that would otherwise be treated monolithically. In the multi-agent workflow tool, the operative structure is a three-level hierarchy over candidate workflows, combined with visualization and pattern-based guidance (Hao et al., 21 Jul 2025). In the CFD surrogate, the operative structure is a compiled traversal-and-staging plan that constrains information flow to bounded local context (Zhang et al., 21 Apr 2026). This suggests a broader methodological resemblance: each system replaces opaque global search or global inference with a staged process whose intermediate states are either visualized for human intervention or compiled for predictable execution.

The contrast between the two systems is nevertheless fundamental. Workflow FlowForge is an interactive client-side web application aimed at practitioners designing LLM-based agent systems, with human-centered evaluation against LangGraph Studio. Rollout FlowForge is a numerical surrogate for PDE and CFD benchmarks, evaluated through RMSE, divergence error, robustness to corruption, and inference latency. One centers on human-in-the-loop design-space exploration, the other on locality-preserving autoregressive prediction. The shared name should therefore be understood as homonymous rather than as indicating a unified technical framework.