Self-Evolving Agentic Workflows via Semantic Decomposition of Complex Tasks

Core Contribution

The paper proposes SeaFlow, a framework in which a coding agent (GPT-5.5) iteratively writes and revises an executable workflow for interactive economic games (the GLEE bargaining, negotiation and persuasion environments). The workflow is a hierarchy of subtasks. Each subtask has an explicit applicability condition (a "semantic" situation) and a solution procedure, which is code, a fitted ML model, or an optional LLM call. Four revision operations are defined: task decomposition, parameter adjustment, model fitting, and procedure redesign. Candidates are retained or reverted after replay or bounded deployment.

The conceptual claim is that naming the situations a step serves, and revising their boundaries, differs from prior workflow-optimization work (AFlow, ADAS, HyEvo, TRACE). Those methods leave the situation implicit. The headline empirical result is that the agent ranked first in the GLEE competition, with final ratings of about 2760, 2762 and 2262. Its decisions are made locally in about 0.1–0.2 ms, and LLM calls end up almost entirely replaced by code.

Methodological Rigor

The paper is largely a descriptive case study, and it is candid about this. The conclusion says "controlled comparisons are needed to quantify performance gains and costs." There is no ablation of semantic decomposition against non-semantic baselines. There is no comparison with AFlow, ADAS or similar methods on the same task. There is no measure of the cost of the revision process itself (GPT-5.5 calls, iterations), and no isolation of any change's effect on the rating. The appendices repeatedly state that the analyses "do not attribute performance gain to edits."

The "ranks first" claim rests on a single competition snapshot and is not accompanied by controlled variance estimates. Rating histories are noisy and non-monotonic (Negotiation drops from 2923 to 1844 before recovering). The "iterations" are commits drawn from a merged index of branches, and the logs are not releasable.

The strongest rigor is in the forensic audit of the development process:

The audit is honest but descriptive. The operation labels are heuristic syntactic rules. The finding that the LLM judge almost vanishes is partly an artifact of the agent's own design decisions. The latency comparison (code in milliseconds versus failed LLM calls over seconds) is trivially expected. The comparison uses five LLM calls, three of which failed.

Potential Impact

The idea of compiling agent reasoning into explicit, testable, conditioned code with LLMs reserved for residual language decisions is useful. It fits a growing "LLM as code-writer, not runtime decider" trend, along with LLM-as-Code and HyEvo, which the authors cite and partly overlap with. The qualitative evidence is interesting: strategies are proposed, shadow-tested, withdrawn and repaired, and revisions follow a replay-first, regression-tested protocol. This could inform practitioners building self-improving coding-agent loops for latency-sensitive or adversarial settings.

Generality beyond GLEE's small, structured, low-dimensional games is untested. Language-heavy tasks, where code cannot replace the LLM, are the real challenge for the approach. The competition-specific nature (a workshop track) and the reliance on a very strong proprietary coding model further limit transfer. The framework is also not formalized beyond Eq. 2. A is simply a coding agent given a prompt protocol, so the "method" is mostly a prompting and engineering pipeline.

Timeliness & Relevance

The paper is timely given interest in agent inference cost, self-evolving agents, and interpretable agent behavior (it was submitted to an interpretability workshop). Code-first agents are a live theme. GLEE is a niche benchmark, but competition-winning results carry some credibility.

Strengths

Limitations

Overall

This is a workshop-level system and case-study paper. It has a credible practical result and a thorough process analysis. It lacks the controlled evidence needed to establish the core methodological claim, so its likely impact is moderate and mostly local to agentic-workflow and game-agent researchers.