CompAgent: A Compositional Agent Framework

• CompAgent is a computational agent that decomposes complex tasks into manageable subtasks using planning, tool selection, and self-correction.
• It is applied across domains such as text-to-image generation, long-horizon context compression, multi-agent control, and wireless network optimization.
• Empirical evaluations show that CompAgent systems enhance fidelity, efficiency, and convergence compared to traditional monolithic approaches.

A "CompAgent" refers to a computational or compositional agent—across multiple research threads—that autonomously decomposes complex tasks into tractable subcomponents, executes and coordinates those components (often using planning, tool use, or self-correction), and, in advanced forms, leverages memory or optimization techniques to improve efficiency and reliability. The term has multiple contextual realizations in the literature: compositional reasoning for text-to-image generation (Wang et al., 2024), context compression for long-horizon agents (Kang et al., 1 Oct 2025), agentic frameworks for wireless network design (Li et al., 27 Jan 2026), and control-theoretic protocols in multi-agent systems (Meng et al., 2013). This article surveys the main instantiations, methodologies, and theoretical principles governing CompAgent systems.

1. Compositional Agents in Multi-modal Generation

CompAgent, as described in "Divide and Conquer: LLMs can Plan and Self-Correct for Compositional Text-to-Image Generation," operationalizes the divide-and-conquer paradigm for compositional text-to-image (T2I) tasks, wherein an LLM agent decomposes a complex prompt into atomic objects, attributes, and inter-object relationships. The agent orchestrates:

• Decomposition: Parsing a prompt into discrete objects, attributes, spatial/non-spatial relationships, and generating scene layouts (bounding boxes).
• Planning and Tool Selection: Routing among a toolkit comprising a multi-concept customization model (for attribute binding), layout-to-image models (for relationship enforcement), and local image editing for post-hoc corrections.
• Verification and Self-Correction: Employing vision-LLMs (e.g., GPT-4V) for attribute verification; invoking object-level edits where failures are detected.

The system achieves substantially improved fidelity in attribute binding and inter-object composition, outperforming prior T2I models by more than 10% across core metrics such as color, shape, and complex scene understanding (T2I-CompBench) (Wang et al., 2024).

2. Context Compression and Long-Horizon Reasoning

In "ACON: Optimizing Context Compression for Long-horizon LLM Agents," the CompAgent paradigm focuses on compressing the working context of long-horizon LLM agents. The central framework, Agent Context Optimization (ACON), is formalized as:

• Compression Problem: For an agent operating over $T$ steps in a POMDP, with history $C = (o_1, a_1, ..., o_T)$, the goal is to derive a compressed context $\tilde C$ minimizing the cumulative token length $C(\tilde C)$ while retaining high terminal task reward $R(s_T)$.
• Optimization: The guideline $G$ for context reduction is iteratively refined via LLM-driven contrastive feedback (comparing failure cases under compression against successes without).
• Distillation: Once an optimized $G^*$ is learned, the compressor is distilled into a compact student model, ensuring low overhead ($95\%$+ retention of performance at $2$--$10\times$ speedup).

Empirically, ACON-augmented CompAgents yield $C = (o_1, a_1, ..., o_T)$0--$C = (o_1, a_1, ..., o_T)$1 reduction in peak memory usage with negligible impact on success, enabling efficient deployment of smaller LLM agents on long-horizon tasks (Kang et al., 1 Oct 2025).

3. Multi-agent Systems with Structural Constraints

"Multi-agent Systems with Compasses" formalizes a kind of CompAgent in continuous-time networked control. Here, each agent holds a "compass"—a shared global orientation—and dynamics are designed to meet generalized tangent-cone conditions:

• Dynamics: For $C = (o_1, a_1, ..., o_T)$2 agents in $C = (o_1, a_1, ..., o_T)$3 with states $C = (o_1, a_1, ..., o_T)$4, the system is

$C = (o_1, a_1, ..., o_T)$5

where $C = (o_1, a_1, ..., o_T)$6 is each agent’s control vector.

• Tangent Cone Protocols: Rather than restricting $C = (o_1, a_1, ..., o_T)$7 to the convex hull of neighbors, it suffices that $C = (o_1, a_1, ..., o_T)$8 belongs to a strict tangent cone based on the supporting hyperrectangle of agent $C = (o_1, a_1, ..., o_T)$9 and its neighbors, facilitated by access to shared reference directions.
• Convergence: Under uniform joint (quasi-)strong connectivity, cooperative networks achieve exponential agreement, while the cooperative–antagonistic extension yields componentwise absolute-value consensus (Meng et al., 2013).

This relaxation expands the admissible dynamics over convex-hull-based consensus, offering accelerated convergence and greater protocol flexibility.

4. Agentic Architectures in Wireless Network Optimization

"ComAgent: Multi-LLM based Agentic AI Empowered Intelligent Wireless Networks" generalizes CompAgent to multi-LLM agentic systems for intent-driven, cross-layer optimization in wireless domains:

• Agentic Cognitive Loop: Four specialized agents each tackle Perception (task and context parsing), Planning (hierarchical decomposition down to solver selection), Action (data and code generation), and Reflection (error and feasibility checking).
• Recursive Decomposition: Each problem is recursively split into subtasks via chain-of-thought and plan-and-solve prompting, then solved via tool or code invocation, with structured memory facilitating cross-agent coordination.
• Self-correction: The reflection loop incorporates compile/runtime error catching, physics-aware constraint validation (e.g., SINR, energy budgets), and triggered code/model revision.
• Results: On beamforming and generic cross-layer tasks, ComAgent architectures demonstrate 100% code execution rates and outperform monolithic LLM solutions on problem formulation (100% vs. 0–56%) and solution rates (72% vs. 24–56%) (Li et al., 27 Jan 2026).

This expert-inspired orchestration is shown to generalize to nontrivial, solver-ready mathematical problem spaces.

5. Evaluation Metrics and Empirical Insights

Across instantiations, CompAgent systems are evaluated on domainspecific, compositional, or agentic benchmarks:

• T2I-CompBench: Attribute binding (BLIP-VQA), spatial/non-spatial relationship AP (UniDet, CLIPScore), and composite scene metrics for text-to-image CompAgent (Wang et al., 2024).
• Long-horizon Agent Tasks: Success rate retention and peak-token reductions in AppWorld, OfficeBench, and QA chains, comparing vanilla and ACON-compressed agents (Kang et al., 1 Oct 2025).
• Wireless Optimization: Problem formulation, code execution, and solved-rate in agentic multi-LLM frameworks versus single-LLM baselines (Li et al., 27 Jan 2026).
• Control-Theoretic Consensus: Exponential rates of agreement and structural conditions required for global convergence in decentralized networks (Meng et al., 2013).

Empirical results consistently indicate the advantages of compositionally structured, self-correcting, and memory-efficient CompAgent systems over monolithic or non-agentic approaches.

6. Limitations and Open Challenges

Notable limitations and future research questions, as identified in the primary sources, include:

• Agentic systems may induce inference or routing overhead due to cross-agent or recurrent LLM prompting, constraining real-time applicability in highly time-sensitive domains (Li et al., 27 Jan 2026).
• Current episodic designs in agentic wireless frameworks lack persistent, event-driven operation, with absence of long-term memory storage for plan templates or failure cases (Li et al., 27 Jan 2026).
• Generative compression (e.g., ACON) can disrupt standard KV-cache use, suggesting future work on hybrid retrieval/summarization techniques for efficient context window management (Kang et al., 1 Oct 2025).
• In compositional T2I CompAgent workflows, LLM-driven decomposition and human-in-the-loop layout adjustments are still required for complex or out-of-distribution prompt structures (Wang et al., 2024).

A plausible implication is that future CompAgent systems will require innovations in persistent, distributed memory, hierarchical agent architectures, dynamic tool orchestration, and end-to-end trainability for robust operation across diverse, real-world domains.

7. Cross-domain Synthesis and Outlook

The CompAgent archetype—whether in vision/language generation, long-horizon planning, agentic network control, or multi-LLM collaboration—converges on several principles: recursive/atomic decomposition of complex inputs, dynamic planning and tool invocation, feedback-driven self-correction, and memory-efficient context management. These design patterns enable robust, compositional reasoning previously unattainable by single-step or monolithic architectures. The progress documented in recent literature points to CompAgent frameworks as foundational in the next generation of autonomous, adaptive AI and multi-agent control systems (Wang et al., 2024, Kang et al., 1 Oct 2025, Li et al., 27 Jan 2026, Meng et al., 2013).