SEW: Self-Evolving Agentic Workflows for Automated Code Generation

1 Introduction

• Complex coding tasks are increasingly challenging for a single LLM instance, motivating multi-agent collaboration.Multi-agent systems decompose programming tasks into subtasks handled by specialized agents with tailored prompts.Source: §1 · p. 1
• Current multi-agent workflows require manually crafting both topology and prompts, limiting adaptability across coding tasks.The paper notes that workflows optimized for machine-learning tasks can differ significantly from those tailored to software development.Source: §1 · p. 1
• SEW automatically generates and optimizes multi-agent workflows, including both workflow topology and each agent’s prompt.Its evolutionary scheme improves both, while the paper also compares five textual representation schemes for interpretation and optimization.Source: §1 · pp. 1–2
• The paper compares BPMN, CoRE, Python, YAML, and pseudo-code as workflow representations for SEW.The comparison considers how well each scheme can be interpreted and optimized by SEW.Source: §1 · p. 2
• The authors report experiments on MBPP, HumanEval+, and LiveCodeBench showing that SEW improves workflow performance through self-evolution.They describe the framework as constructing workflows from scratch based solely on task descriptions.Source: §1 · p. 2

2.1 Workflow Representations in Agents

• BPMN, CoRE, Python, YAML, and pseudo-code are described as workflow representation schemes used in agentic systems.BPMN depicts execution order, while CoRE combines natural-language programming, pseudo-code, and flow-based programming.Source: §2.1 · p. 2

2.2 Self-Evolving Agents

• Prior work includes prompt optimizers and methods for evolving agentic workflows, while SEW jointly optimizes agent prompts and workflows across representation schemes.Examples include APE’s candidate search, Promptbreeder’s prompt mutation, MIPRO’s instruction and example refinement, AFlow’s Monte Carlo Tree Search, and EvoFlow’s automated workflow search.Source: §2.2 · p. 2

3 SEW

• The task is to generate executable code from a textual coding problem through collaboration among LLM agents.Each agent processes a textual prompt and produces a textual output; a workflow is a sequence of agents.Source: §3 · p. 2
• SEW’s three modules are workflow generation, workflow evolution, and agent evolution.It first generates an initial workflow from the task description, reconstructs it through workflow evolution, then equips each agent with a more sophisticated prompt.Source: §3 · p. 3, Figure 2 · p. 3
• An LLM generates default workflows from the task description and a template expressed in one of five representation schemes.The schemes are BPMN, CoRE, Python, YAML, and pseudo-code.Source: §3 · p. 3
• Workflow evolution applies a Direct Evolution operator to a textual workflow using a mutation prompt to produce an evolved workflow.SEW evaluates representation schemes using Logical Successful Rate for workflow validity and Generation Successful Rate for executable Python output.Source: §3 · pp. 3–4
• Direct Evolution modifies an agent’s prompt, whereas Hyper Evolution first modifies the mutation prompt and then applies it to the agent.The framework combines workflow and agent evolution to generate workflow variants for code generation.Source: §3 · pp. 4–5

• A useful representation must capture workflow structure and semantics while remaining interpretable to LLMs for downstream modification.Source: §4 · p. 5
• SEW compares BPMN, CoRE, Python, YAML, and pseudo-code to identify suitable workflow representations for LLM understanding and execution during workflow evolution.Source: §4 · pp. 5–6
• Figure 4 contrasts BPMN and CoRE representations of one software-development pipeline, using task nodes for stages and sequence flows for dependencies; both are intended to perform the same function.Source: Figure 4 · p. 5, §4 · p. 6
• BPMN emphasizes task order and dependencies, while CoRE combines natural-language programming, pseudo-code, and flow-based programming and supports executable, LLM-interpretable workflows.Source: §4 · p. 5
• Python offers familiar, flexible programmatic integration; YAML supports readable hierarchical configuration, while pseudo-code bridges natural language and formal code for human and machine readability.Source: §4 · p. 5

5.1 Dataset

• LCB testing uses 300 code-generation samples after 100 randomly sampled validation examples; MBPP and HumanEval+ follow AFlow's data split.The LCB set is the release_v1 code-generation subset.Source: §5.1 · p. 6, §5.2 · p. 6

5.2 Baselines

• Baselines include two backbone models, Chain-of-Thought, ADAS, AFlow, and PromptBreeder across Claude Opus 4.6 and GPT-5.3.The evaluation covers HumanEval+, MBPP, and LCB.Source: §5.2 · p. 6

5.3 Experimental Setup

• The authors select task parsing and code rewriting workflows for analysis because variants built on them often improve performance over other generated workflows.They search workflows represented in BPMN, CoRE, Python, YAML, and pseudo-code.Source: §5.3 · p. 6

5.4 Main Results

• SEW outperforms both backbone models, Chain-of-Thought, ADAS, and AFlow under the same settings; Claude Opus 4.6 generally outperforms GPT-5.3 across the three datasets.The authors report the Claude-versus-GPT pattern on MBPP and LCB.Source: §5.4 · p. 7

5.5 Analysis

• 91.5% is BPMN's highest LSR, but CoRE achieves the best GSR and is identified as the most effective workflow representation.The authors describe CoRE as combining natural-language, pseudo-code, and flow programming.Source: §5.5 · p. 7
• SEW’s workflow evolution produces code-rewriting workflows that consistently outperform the Claude Opus 4.6 backbone across three datasets; agent evolution improves task-parsing performance by 9.5% on LCB.The authors attribute the initial improvement to novel workflow topologies and further gains to high-quality prompts generated for each agent.Source: §5.5 · p. 8
• HE shows lower variance than DE, while second-order DE reaches around 0.57–0.60 pass@10 for Code Rewriting; the authors frame the choice as robustness versus peak performance.They describe lower-order DE and zero-order HE as more robust, and second-order DE and first-order HE as suited to maximizing performance with some variability.Source: §5.5 · p. 8, Figure 5 · p. 8

6 Conclusion

• Across three coding benchmarks, self-evolved workflows outperform single-agent baselines even when those baselines use improved prompts.The conclusion also reports that higher-order evolution strategies further improve workflow effectiveness.Source: §6 · p. 8
• Higher-order evolution strategies further improve workflow effectiveness by refining task execution dynamics.The conclusion describes this as an additional improvement to workflow effectiveness.Source: §6 · p. 8
• CoRE is the most effective of the tested workflow representation schemes, balancing logical correctness and execution success.The conclusion presents this as the result of comparing different representation schemes.Source: §6 · p. 8
• SEW offers a direction for reducing reliance on manual workflow design and prompt engineering while improving adaptability and efficiency.These are the conclusion's stated implications for agentic workflow optimization.Source: §6 · p. 9

Limitations

• SEW’s generalization to other AI-driven tasks, such as reasoning or planning, remains unknown.The authors frame this as an open boundary on its demonstrated applicability.Source: Limitations · p. 9
• Some generated workflows that are logically sound fail to produce executable outputs.The authors identify a need for more robust validation mechanisms in response.Source: Limitations · p. 9
• SEW’s effectiveness depends on the capabilities of the underlying LLM.The authors note that future LLM architecture advancements could significantly affect its performance.Source: Limitations · p. 9

A.1 Workflow Representation Schemes

• The five workflow representation schemes are BPMN, CoRE, Python, YAML, and pseudo-code.The appendix identifies BPMN and CoRE as scripts and presents Python, YAML, and pseudo-code versions of the same workflow.Source: §A.1 · p. 10
• The Python and YAML examples encode named agents as ordered steps with specified arguments and outputs.The examples include task parsing, task refinement, code generation, code review, and code refinement agents.Source: §A.1 · p. 10
• The pseudo-code example expresses the workflow as agent calls that pass outputs such as parsed_task and generated_code to later steps.It ends with code refinement using the refined task and review comments.Source: §A.1 · p. 10

A.2 Prompt Corpus

• The prompt corpus describes LiveCodeBench, HumanEval+, and MBPP as code-generation tasks with different problem and evaluation formats.LiveCodeBench uses unseen test cases, HumanEval+ contains 164 handcrafted problems, and MBPP has approximately 1,000 crowd-sourced problems.Source: §A.2 · p. 11
• A template workflow guides LLMs in generating default workflows, while agent prompts specify roles, objectives, inputs, and outputs.The default workflow prompt also asks for workflow steps, sequence, and agent responsibilities based on a workflow template and dataset description.Source: §A.2 · p. 11
• The evolutionary prompt set includes mutation, hyper-mutation, and thinking-style prompts, which the paper locates in PromptBreeder’s appendix..Source: §A.2 · p. 12

A.2.1 Workflows without agent evolution module

• Before agent evolution, the Task Parsing Workflow passes a task summary to code generation, while the Code Rewriting Workflow adds review and rewriting stages.The appendix notes that reviewer and rewriting agents may form a loop.Source: §A.2.1 · p. 12

A.2.2 Workflows improved by agent evolution module

• An evolved code example adds function extraction, error handling, and invalid-input feedback to a square-number program.The example also discusses clearer code and handling edge cases.Source: §A.2.2 · p. 13
• The evolved mutator example asks for unconventional combinations and applications, while reviewer suggestions call for flexible inputs, randomness, and documentation.The prompt also invites visual or narrative outputs.Source: §A.2.2 · pp. 13–15
• A separate example generates random words and uses them to create a scatter plot as visual art.The accompanying code accepts the requested number of words and displays the resulting plot.Source: §A.2.2 · pp. 15–16

A.3 Examples of Failed Python Workflows

• One failed workflow ends with a reviewer rather than a coding agent, so its final output cannot solve the coding task.Source: §A.3 · p. 16
• The illustrated pipeline parses and refines the task, generates code, then passes the task and generated code to a reviewer.The example sequence names task parsing, task refinement, code generation, and code review agents.Source: §A.3 · p. 16
• Another workflow routes generated code to a fortune-telling agent expecting refined_code, leaving that agent unable to fetch its input.Source: §A.3 · p. 16

A.4 Token costs

• The comparison covers backbone-only GPT-4o mini and SEW-generated Task Parsing and Code Rewriting workflows.Source: §A.4 · p. 16
• LiveCodeBench has higher per-instance token costs than HumanEval+ and MBPP.Source: §A.4 · p. 16
• Both workflows cost more tokens than GPT-4o mini, with Code Rewriting especially costly because code generation may take multiple reviewer rounds.Source: §A.4 · p. 16