Research map / Multi-agent systems and games / LLM agents
Flow: Modularized Agentic Workflow Automation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Flow, an LLM multi-agent framework that represents workflows as activity-on-vertex graphs and dynamically refines them during execution, favoring parallelism and modularity.
Problem
Existing LLM multi-agent frameworks use static or sequential workflows that cannot adapt efficiently to unforeseen failures during task execution.
Approach
Model the workflow as an AOV DAG stored as a JSON-like dictionary, generate several candidate graphs with an LLM, and select the one with highest average parallelism and lowest dependency-degree standard deviation; the LLM global inspector updates the graph locally based on subtask outputs.
Key results
- Flow achieves 100% overall success on Gobang and LaTeX Beamer tasks and 80% on website design, averaging 93% success versus 66.7% for AutoGen, 71% for MetaGPT, and 48.67% for CAMEL (GPT-4o-mini).
- Flow human rating averages 3.54/4 across tasks versus 2.63 for AutoGen, 1.60 for MetaGPT, and 2.12 for CAMEL.
- With GPT-3.5-Turbo, Flow scores 87 on Gobang, 80 on website design, and 67 on LaTeX Beamer, outperforming baselines on the first two.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Flow instead of MetaGPT | multi-agent framework · Gobang Game Development | Overall Score (%): 73 -> 100 | Helped |
| Use Flow instead of AutoGen | multi-agent framework · LaTeX Beamer Writing | Overall Score (%): 67 -> 100 | Helped |
| Use Flow instead of MetaGPT | multi-agent framework · Website Design | Overall Score (%): 80 -> 80 | No clear effect |
| Use Flow instead of CAMEL | multi-agent framework · Gobang Game Development | Overall Score (%): 67 -> 87 | Helped |
| Use Flow instead of MetaGPT | multi-agent framework · Website Design | Overall Score (%): 67 -> 80 | Helped |
| Use Flow instead of CAMEL | multi-agent framework · LaTeX Beamer Writing | Overall Score (%): 53 -> 67 | Helped |
| Add dynamic workflow updates (ablation) | Flow · Gobang Game Development (GPT-3.5-Turbo) | Execution time: 26.12 ± 11.35 -> 33.57 ± 12.46 | Hurt |
| Add dynamic workflow updates (ablation) | Flow · Website Design (GPT-3.5-Turbo) | Execution time: 23.46 ± 10.84 -> 34.23 ± 13.12 | Hurt |
| Add dynamic workflow updates (ablation) | Flow · LaTeX Beamer Writing (GPT-3.5-Turbo) | Execution time: 18.34 ± 9.73 -> 24.12 ± 10.89 | Hurt |
Surprising findings
- Dynamic updates increase execution time (e.g., Gobang 26.12s to 33.57s with GPT-3.5-Turbo), a trade-off the authors acknowledge.
- Flow scored 0% on the page-limit criterion for LaTeX Beamer with GPT-3.5-Turbo, the same as the baselines.
Limitations
- Evaluation uses only three self-designed coding and writing tasks with five trials each, so statistical power is limited.
- Dynamic updates add API calls and execution overhead.
- Human ratings come from 50 participants without reported inter-rater agreement details.
Experimental setup
Benchmarks: Gobang game development LaTeX Beamer writing Website design
Metrics: success rate human rating execution time changed ratio
Baselines: AutoGen CAMEL MetaGPT
Builds on: AutoGen MetaGPT CAMEL ReAct Data Interpreter
Closest papers by meaning
- GraphFlow: A Graph-Based Workflow Management for Efficient LLM-Agent Serving (ICML 2026 · similarity 0.91)
- Benchmarking Agentic Workflow Generation (ICLR 2025 · similarity 0.90)
- Optimize Once, Execute Fast: Latency-Aware Multi-Agent Workflow Learning for Recurrent Queries (NeurIPS 2026 · similarity 0.90)
- FlowSteer: Towards Agents Designing Agentic Workflows via Reinforced Progressive Canvas Editing (NeurIPS 2026 · similarity 0.89)
- DyFlow: Dynamic Workflow Framework for Agentic Reasoning (NeurIPS 2025 · similarity 0.88)
- Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.