In Brief
A no-code platform lets non-programmers design, run, and inspect complex multi-agent workflows while reliably handling nested feedback loops and preserving per-iteration state.
ON THIS PAGE
Key Findings
DevAll (ChatDev 2.0) turns visual YAML specs into runnable multi-agent graphs so users can compose agents, tools, and data flows without writing orchestration code. Semantic edges separate how data moves from when agents run, letting the system route information and trigger work independently. A cycle-aware scheduler called CADET condenses tightly linked cyclic regions and executes them with iteration-level state, so complex feedback patterns run correctly. Across three domains—data visualization, deep research reports, and software generation—DevAll reproduced core workflows on a single platform with competitive results. human-in-the-loop hooks
Key Data
17 built-in node types capture common agent and tool roles in the visual canvas
23 representative domains used for evaluation: scientific visualization, deep research report generation, and software development
31 cycle-aware scheduler (CADET) that condenses cyclic regions and preserves iteration-level state to support arbitrary nested feedback loops
Why It Matters
Product and platform engineers building AI-powered pipelines can prototype complex agent teams faster without writing orchestration code. Technical leads and researchers evaluating multi-agent approaches can use DevAll to compare workflows, inspect traces, and add human feedback loops without heavy engineering overhead. multi-agent approaches
Avoid common pitfallsLearn what failures to watch for
Key Figures
Fig 1: Figure 1: Users drag heterogeneous components onto the canvas and connect them into executable multi-agent systems. DevAll supports cross-domain artifact generation, including charts, research reports, and so on.
Fig 2: Figure 2: Overview of DevAll. The upper part shows the user interface for graph authoring, execution launching, and result inspection; the lower part presents the underlying engine, which includes the MAS Compilation Engine that compiles declarative specifications into executable graphs, and the CADET runtime that schedules cycle-aware execution over semantic edges to produce final results along with intermediate traces and artifacts.
Fig 3: Figure 3: CADET execution over MAS graphs. CADET contracts cyclic regions into condensed nodes for layer-wise scheduling, and executes the corresponding SCC subgraphs through a scoped iterative process.
Fig 4: Figure 4: Qualitative comparison of generated visualization artifacts on a representative MatPlotBench case.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
No-code here lowers implementation work but does not automate workflow design—users still must decide agent roles, prompts, and control conditions. The current component set and graph abstraction may not cover every domain-specific interaction, so some integrations may need custom extensions. Scalability, long-term maintenance of large visual graphs, and usability for novice authors require further study and tooling (templates, lifecycle management, user tests). lifecycle management
Methodology & More
DevAll provides a visual, no-code environment where users author multi-agent workflows as YAML-based graphs. Nodes represent agents, tools, or data operations and edges are semantic: each edge defines when it should fire, how to merge or transform data, and how it affects downstream scheduling. By separating data flow from control triggers, the platform supports richer interaction patterns than fixed, linear workflows and avoids manual state plumbing between steps. At runtime, the CADET scheduler handles arbitrary cycles by collapsing tightly coupled cyclic regions into condensed units (think: treat a loop of agents as one block) and running those blocks with scoped, iteration-aware execution to preserve per-iteration state. The team reimplemented three domain workflows—chart generation, deep research report creation, and software generation—using only declarative specs and a common backbone model, showing that a single no-code platform can reproduce diverse specialist systems. DevAll is open-sourced and includes trace inspection and human-in-the-loop hooks, making it practical for prototyping, debugging, and evaluating multi-agent designs without bespoke orchestration code. Planning Pattern
Need expert guidance?We can help implement this
Credibility Assessment:
ArXiv preprint with multiple authors; some recognizable names but low listed h-indexes. Overall fits a solid/recognized researcher level rather than top-tier venue.