Why It Matters
For agent infrastructure, this is a clear pre-runtime map. Each module implies state management, prompt/schema control, tool execution, evaluation hooks, and failure modes.
Problem
- Traditional autonomous agents were often trained with limited knowledge in isolated environments.
- LLMs made open-domain autonomous agents plausible, but early systems were proposed independently with inconsistent terminology.
- Research on profiles, memory, planning, tools, and evaluation was fragmented across many agent papers.
- The field needed a taxonomy that connected construction choices to applications and evaluation.
Key gap: The survey asks how to systematically summarize LLM-based autonomous agent construction, applications, evaluation strategies, and challenges.
Method
Survey LLM autonomous agents through three linked taxonomies: construction, applications, and evaluation, with challenges derived from the gaps in those categories.
- Define the scope around LLM-based autonomous agents rather than general chat models.
- Propose a unified construction framework with profile, memory, planning, and action modules.
- Survey memory structures, formats, and operations, including short/long-term memory and database-like stores.
- Survey planning methods with and without feedback, including external planners and ReAct-style environment feedback.
- Survey capability acquisition through fine-tuning, prompt engineering, and mechanism engineering.
- Classify applications in social science, natural science, and engineering.
- Classify evaluation into subjective and objective strategies and summarize unresolved challenges.
Claims
C1 · paper-supported
The survey proposes a unified construction framework with profiling, memory, planning, and action modules.
Scope: Section 2.1.
- E-framework: The paper's unified framework for autonomous agents is composed of profiling, memory, planning, and action modules.
C2 · paper-supported
The profile module can be handcrafted, generated by an LLM, or aligned from real-world datasets.
Scope: Section 2.1.1.
- E-profile: The survey describes handcrafted profiles, LLM-generated profiles, and dataset-aligned profiles as common ways to specify agent roles.
C3 · paper-supported
The memory module is analyzed through structures, formats, and operations such as reading, writing, and reflection.
Scope: Section 2.1.2.
- E-memory: Memory is discussed through short/long-term structures, formats including natural language, embeddings, databases, and structured lists, and operations including reading, writing, and reflection.
C4 · paper-supported
Planning methods are organized into no-feedback strategies, external-planner strategies, and feedback-driven planning from environments, humans, or models.
Scope: Section 2.1.3.
- E-planning: Planning methods include single-path and multi-path reasoning, external planners such as PDDL-based approaches, and feedback-driven planning such as ReAct-style thought-action-observation loops.
C5 · paper-supported
The survey groups LLM-agent applications into social science, natural science, and engineering domains.
Scope: Section 3.
- E-applications: The application survey covers social science, natural science, and engineering, including software engineering, robotics/embodied AI, chemistry, and social simulation examples.
C6 · paper-supported
Agent evaluation is split into subjective evaluation by humans or proxies and objective evaluation using metrics, protocols, and benchmarks.
Scope: Section 4.
- E-evaluation: The paper distinguishes subjective evaluation, such as human annotation and Turing-test-style comparison, from objective evaluation using quantitative metrics, protocols, and benchmarks.
C7 · inference
The profile-memory-planning-action decomposition is directly usable as a review schema for agent papers because it separates identity, state, control, and effects.
Scope: Application to agent-readings review structure.
- E-schema-implication: Profile, memory, planning, and action correspond to identity/configuration, persistent state, control policy, and side-effect surface, which are core schema axes for structured reviews.
Agent Infrastructure Takeaways
- This survey gives a clean architecture checklist for reviewing agent papers: profile, memory, planning, action, capability acquisition, and evaluation.
- Memory formats matter operationally. Natural-language, embedding, database, and structured-list stores expose different retrieval, audit, and mutation semantics.
- Planning with feedback is the critical line between static prompt decomposition and real agent control loops.
- Profiles should be treated as configuration with provenance, not harmless prompt text, especially in multi-agent simulations.
- Evaluation needs subjective and objective signals, but neither is sufficient alone for production-grade confidence.
- The construction taxonomy is directly useful for this project's review database because it separates mechanisms that Markdown-only summaries tend to blur.
Caveats
- The arXiv paper has later revisions; this review preserves the reading-list year while noting the source is actively maintained.
- As a survey, it synthesizes primary papers rather than producing new experimental results.
- The taxonomy predates many later agent-runtime, safety, and benchmark papers.
- The application survey includes early-stage prototypes and research demonstrations, so maturity varies widely.