ChatDev 2.0: A No-Code Multi-Agent Platform for Developing Everything
Yufan Dang, Shu Yao, Bowen Lai, Chenting Xu, Ruijie Shi, Wai-Shing Leung, Huatao Li, Chen Qian, et al.
cs.AI · cs.CL · cs.MA
Quick verdict
No-code MAS platform with cycle-aware CADET scheduler; reproduces three workflows at competitive performance, but evaluation lacks statistical rigor.
Summary
The paper presents ChatDev 2.0 (DevAll), a no-code platform for building, executing, and inspecting LLM-based multi-agent systems. Users assemble graphs on a visual canvas from seven node types (Agent, Human, Python, Subgraph, Literal, Passthrough, LoopCounter); a compilation engine turns a YAML specification into an executable graph whose 'semantic edges' decouple data-flow policies from control-flow activation. The CADET execution engine handles cyclic graphs by decomposing them into strongly connected components (Tarjan 1971), condensing them into a schedulable DAG, and running each cyclic component as a scoped iteration whose back edges are deferred to the next round, bounded by an iteration limit. To test expressiveness, the authors re-express three specialized systems — CoDA (data visualization), Enterprise Deep Research, and ChatDev 1.0 (software generation) — as declarative graphs with a standardized GPT-4o backbone. DevAll scores 0.7950 versus CoDA's 0.7130 on MatPlotBench, 0.3319 versus EDR's 0.3500 on DeepResearchBench, and 0.6509 versus ChatDev 1.0's 0.6574 on SRDD. Scheduling overhead is small (16.23 ms compilation at 16 SCCs). Main limitations: no statistical tests or variance reporting despite judge-based metrics, unclear baseline re-run conditions, no ablation of CADET against alternative loop mechanisms, and no user study of authoring efficiency.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A declarative graph with semantic edges (c_e, φ_e, ψ_e) can express heterogeneous MAS, including dynamic and cyclic interactions, without any task-specific orchestration code.
- CADET executes arbitrary cyclic MAS graphs by condensing strongly connected components into a DAG and running scoped iterations, with millisecond-level scheduling overhead.
- DevAll reproduces CoDA, EDR, and ChatDev 1.0 at competitive overall scores (0.7950 vs 0.7130; 0.3319 vs 0.3500; 0.6509 vs 0.6574) on a unified GPT-4o backbone.
- No-code authoring reduces specification effort to 4.6–54.4% of the reference systems' source size (Table 5), at 84–258 estimated canvas operations (Table 2).
Strengths
Clean expressiveness abstraction
The semantic edge tuple e = (u, v, c_e, φ_e, ψ_e) (§3.1) explicitly separates activation conditions, data-flow policies, and scheduling control, giving a principled answer to the workflow-centric limits the paper documents for Dify Loop/Iteration, Langflow Loop, and Flowise iteration subflows (§2).
CADET is grounded and cheap
Cyclic handling rests on standard SCC condensation (Tarjan 1971; Bang-Jensen & Gutin 2009) with full pseudocode (Algorithms 1–2), and overhead is measured properly: p50/p95 over 1,000 runs with five warm-ups, reaching only 16.23 ms compilation and 91.96 μs planning at 16 SCCs (Table 3).
Cross-domain validation on a standardized backbone
Authoring cost is quantified, not just asserted
Candid limitations section
Weaknesses
No inferential statistics anywhere
Baseline provenance and scale unstated
Central differentiator never tested head-to-head
Post-hoc causal attributions without ablation
Execution semantics under-specified at the edges
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
ChatDev 2.0 ('DevAll') is a systems paper introducing a no-code platform for LLM-based multi-agent systems. It makes three claims: (1) a declarative graph with 'semantic edges' — the tuple e = (u, v, c_e, φ_e, ψ_e) of activation condition, data-flow policy, and control-flow policy (§3.1) — expresses heterogeneous MAS, including dynamic and cyclic interaction, without task-specific orchestration code; (2) the CADET scheduler executes graphs with arbitrary cycles by condensing strongly connected components (Tarjan 1971) into a DAG and running each cyclic SCC as a scoped iteration with deferred back edges (§3.2, Algorithms 1–2); (3) three specialized systems — CoDA, Enterprise Deep Research (EDR), and ChatDev 1.0 — are reproduced on the platform at competitive performance (§4, Table 1). This is infrastructure work, and the paper is appropriately modest: the stated goal is 'not to surpass domain-specific systems' (§4.1).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ChatDev 2.0 (DevAll)?
An open-source no-code platform from the OpenBMB/ChatDev team (Tsinghua, SJTU and collaborators) for building LLM-based multi-agent systems: users drag seven node types onto a visual canvas, connect them with 'semantic edges', and the platform compiles a YAML specification into an executable graph with monitoring and human-in-the-loop support.
How does CADET handle cycles in agent graphs?
It finds strongly connected components with Tarjan's algorithm, contracts them into a DAG for global topological scheduling, then executes each cyclic component as a bounded iteration: back edges are deferred to the next round until an exit edge fires or an iteration limit stops the loop.
Does DevAll outperform specialized multi-agent systems?
No — the goal is matching them. On a shared GPT-4o backbone it scores 0.7950 vs CoDA's 0.7130 on MatPlotBench (higher), 0.3319 vs Enterprise Deep Research's 0.3500 on DeepResearchBench and 0.6509 vs ChatDev 1.0's 0.6574 on SRDD (both slightly lower) — competitive overall, per the paper's own framing.
Where can I get the code and workflows?
The paper points to https://github.com/OpenBMB/ChatDev; workflows are distributed as YAML specifications, which the platform treats as its canonical on-disk representation for sharing and reuse.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–5 are caption-only in the extraction; the UI screenshots and qualitative artifacts (Figs. 4–5) could not be visually assessed, so interface quality and output comparability claims rest on captions and tables alone.
- The supplied metadata lists 8 authors, while the paper text includes Zhiyuan Liu as a ninth (corresponding) author; this review follows the paper text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.