Self-Evolving Agentic Workflows via Semantic Decomposition of Complex Tasks

Zichuan Fu; Junjia Qi; Xian Wu; Wenlin Zhang; Yimin Deng; Kaifeng Guo; Xiangyu Zhao

cs.AINov 1, 2026

Self-Evolving Agentic Workflows via Semantic Decomposition of Complex Tasks

Authors: Zichuan Fu, Junjia Qi, Xian Wu, Wenlin Zhang, Yimin Deng, Kaifeng Guo, Xiangyu Zhao

Organizations: City University of Hong Kong · Tencent Jarvis Lab

Abstract

Agentic workflows organize complex tasks into reusable procedures. However, invoking an LLM at every step incurs repeated inference costs and delays. We propose a framework for self-evolving agentic workflows through semantic task decomposition. Each subtask has explicit applicability conditions and a procedure that executes code or invokes a model as needed. An agent uses execution feedback to refine the decomposition and update these procedures. Experiments in the GLEE competition show the framework on bargaining, negotiation and persuasion. Most observed final-decision procedures are implemented in code. Code is available at https://github.com/Applied-Machine-Learning-Lab/NeurIPS2026_SeaFlow.

Figures & tables

Panel of Figure 1 — Figure 1: (a) Bargaining decision paths at revision a1875a0 ; leaves return plan parameters. (b) The agent’s workflow revision cycle, from execution feedback to a retained workflow.

Measure	Bargaining	Negotiation	Persuasion
Official rating $\uparrow$	2760.27	2761.67	2261.58
Mean branches per strategy node	2.13	2.25	2.24
Mean strategy judgments per action	37.61	22.95	7.42
Mean local decision time (ms)	0.118	0.198	0.122

Table 1: Competition ratings and mean statistics of the final workflows (revision a1875a0 ).

Figure 2: Competition rating history and strategy changes on the same 725-iteration axis. Top: recorded deployment segments, one rating per plotted iteration. Bottom: representative strategy changes, with iteration numbers and operation markers.

Panel of Figure 3 — Figure 3: (a) Shares of designed decision modules by executor across iterations. (b) Decision time per executor against strategy judgments per decision (log scale). Crosses mark failed LLM calls.

Figure 4: Share of non-message actions with a valid LLM-judge response per iteration, by task. Triangles mark responses that change the action. Dotted lines mark Persuasion leaving the judge.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Execution order of the final workflows for one move. Within a box, rows are checked from top to bottom and indentation marks nested checks; dashed boxes are optional model calls.

Figure 6: Situations and procedures of the final workflows on replayed decision states. Each row is one situation; its bar gives the share of states, colored by the second-level split, and the right column lists the procedures chosen within that situation.

Panel of Figure 7 — Figure 7: An actual Bargaining revision. Left: the acceptance branch introduced at position 236, with all its conditions and the fallback to the existing procedure. Right: the motivating situation and expected behavior documented in the same commit’s tests. Earlier branches retain precedence. The diagram shows a local code change, not the entire workflow.

Figure 8: Revision activity per 25 iterations. (a) Commits on the main line and on experimental branches. (b) Added, modified and removed workflow functions. (c) Changed code units by revision operation.

Figure 9: Changed code units per 25 iterations by game and revision operation. Shared modules hold code and learned artifacts used by more than one game.

Figure 10: Evidence named in commit messages across iterations. Curves give the smoothed share of iterations with each mark. Ticks below mark the individual commits.

Policy	Steps
Persuasion seller with a low-quality product	1 skip $\rightarrow$ 22 recommend $\rightarrow$ 29 skip $\rightarrow$ 31 recommend until the first sale
Bargaining plateau exit	548 shadow $\rightarrow$ 552 live $\rightarrow$ 564 reverted $\rightarrow$ 579 live with a guard $\rightarrow$ 585 split into finite and open horizons
Bargaining Rubinstein responder	553 shadow $\rightarrow$ 556 live $\rightarrow$ 557 shadow $\rightarrow$ 559 live on a narrower band $\rightarrow$ 560 shadow

Appendix

Table 2: Policies that change direction during development. Each step gives the iteration and the behavior that follows it.

Figure 11: Regression tests during development. (a) Test functions on the main line. (b) Share of strategy commits per 25 iterations that also change tests.

Iteration	Problem	Revision	Test checks that
97	Message wording misread	Refine conditions	a current recommendation outweighs earlier skip wording
110	Opponent style misjudged	Update the estimate	recent stalls, not an early concession, set the style
285	Low-payoff deals accepted	Refine conditions	a non-final offer at the bottom of the field is not accepted
608	Same message rejected again and again	Refine conditions and change the procedure	a rejected counter message is not sent again
676	Opponent misjudged	Update the estimate and refine conditions	the current opponent’s revealed outcomes override the global prior

Appendix

Table 3: Diagnosed problems, the revisions they trigger and the regression tests that record them.

Competition snapshot
Measure	Bargaining	Negotiation	Persuasion
Official rating $\uparrow$	2760.27	2761.67	2261.58
Cumulative games	8,256	27,226	7,435
Fixed-workflow behavior sample
Games with action records	512	199	401
Recorded actions	2,344	1,912	7,993

Appendix

Table 4: Competition results, fixed-workflow samples, and examined strategies. The September 1 snapshot contains 42,917 cumulative games across the three environments. Fixed-workflow statistics describe a separate analysis sample of 1,112 games and 12,249 action records, including messages; they are not additional games to add to the competition total. Ratings are snapshot values, not averages over the fixed batches.

Condition	Eligible	Executed	Recorded action
Opening imitation	63	63	offer $55\%$ self-share
Reservation lock	451	382	reject below $60\%/47\%$
Low-share predictor gate	8	8	reject

Appendix

Table 5: Condition-to-action checks on fixed Bargaining records. Eligibility does not imply execution when an earlier condition takes precedence.

Figure 12: Final decision labels in fixed-workflow records. Grey segments include 31 unclassified Bargaining records and nine unresolved Negotiation records; eight Bargaining records use a predictor. Runtime LLM judging is disabled in all three deployments. Label definitions are environment-specific; detailed counts and bounds are given in the text.

Figure 13: Recorded LLM decisions across two commit windows. (a) Valid model selections as a percentage of recorded non-message actions. (b) Selections that change the program’s proposed action, using the same denominator. Labels include raw counts. Both axes are linear but use different ranges; the plotted events do not count every attempted call.

Agentic Workflows

Agentic workflows coordinate AI agents, tools, and humans across multi-step tasks, emphasizing reliable execution, efficiency, and oversight.

See all 224 papers

HyEvo: Self-Evolving Hybrid Agentic Workflows for Efficient Reasoning
Mar 20, 2026Beibei Xu, Yutong Ye, Chuyun Shen +3Agentic Workflows LLM Agent Workflow Optimization
East China Normal University · Beihang University · Shanghai University of International Business and Economics +1
FlowScout: From Execution Feedback to Reliable Tool-Using Agent Workflows
Aug 10, 2026Shuo Hao, You Lu, Bihuan Chen +1Agentic Workflows LLM Agent Workflow Optimization
College of Computer Science and Artificial Intelligence, Fudan University, China
RSIFlow: Adaptive Self-Improvement Workflows with an Evolving Meta-Agent
Oct 15, 2026Huan Zhang, Mingju Chen, Dongxu Zhou +4LLM Agent Self-Improvement Self-Evolving Agents
Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing, School of Artificial Intelligence, Beihang University · Tsinghua University · Beijing Academy of Artificial Intelligence (BAAI)

Mar 20, 2026J/K move · Enter open · S save

HyEvo: Self-Evolving Hybrid Agentic Workflows for Efficient Reasoning

Beibei Xu, Yutong Ye, Chuyun Shen, Yingbo Zhou+2

East China Normal University · Beihang University · Shanghai University of International Business and Economics+1

Agentic Workflows LLM Agent Workflow Optimization Efficient Language Model Reasoning LLM-Based Program Synthesis

Open paper AskPDF arXiv

Although agentic workflows have demonstrated strong potential for solving complex tasks, existing automated generation methods remain inefficient and underperform, as they rely on predefined operator libraries and homogeneous LLM-only workflows in which all task-level computation is performed through probabilistic inference. To address these limitations, we propose HyEvo, an automated workflow-generation framework that leverages heterogeneous atomic synthesis. HyEvo integrates probabilistic LLM nodes for semantic reasoning with deterministic code nodes for rule-based execution, offloading predictable operations from LLM inference and reducing inference cost and execution latency. To efficiently navigate the hybrid search space, HyEvo employs an LLM-driven multi-island evolutionary strategy with a reflect-then-generate mechanism, iteratively refining both workflow topology and node logic via execution feedback. Comprehensive experiments show that HyEvo consistently outperforms existing methods across diverse reasoning and coding benchmarks, while reducing inference cost and execution latency by up to 19

\times

and 16

\times

, respectively, compared to the state-of-the-art open-source baseline.

Abstract

Agentic Workflows

Explore similar work

HyEvo: Self-Evolving Hybrid Agentic Workflows for Efficient Reasoning

FlowScout: From Execution Feedback to Reliable Tool-Using Agent Workflows

RSIFlow: Adaptive Self-Improvement Workflows with an Evolving Meta-Agent