Agentic Large Language Models, a survey

In plain words

Most language models only answer when asked: you send a prompt and get text back. This survey looks at the newer kind of model that is set up to act. The authors call it an agentic LLM: a model that works out how to reach a goal, takes actions in the world such as calling tools or controlling a robot, and interacts with other agents.

The authors sort the research into those three jobs. Reasoning work helps a model think through steps, reflect on its own answers and look things up. Acting work connects the model to tools, robots and real tasks, such as assistants for medicine and trading. Interacting work puts several models into a shared setting to see how they cooperate, negotiate and form social norms.

The paper also argues that this matters for training. When an agent acts and receives feedback, the data it produces could be used to train models later, which may help where ordinary training data no longer improves results. The paper closes with open problems, mainly hallucination, safety and responsibility. It is a map of other people's work, not a set of new experiments.

TL;DR

• The paper surveys agentic LLMs, defined as models that reason, act and interact §1.1.
• It groups the literature into reasoning §2, acting §3 and interacting §4, with one summary table per group tab. 1 tab. 2 tab. 3.
• The authors argue that agent interaction can generate new training data, which may help when pretraining data stops improving §1.1 §5.1.
• It reports no experiments of its own. Its claims summarise cited work and were not re-checked by this card.
• Use it as a map to the primary papers. Do not take numbers from it.

Problem

The authors list four limits of plain LLMs: prompt engineering is manual and fragile, models hallucinate, multi-step reasoning is weak, and training data is finite §1.4. They also say that gains from scaling data are plateauing §1.1 §1.4. A model that only predicts text has no way to look up new facts, act on the result or collect new experience §1.1. The survey assumes the usual pipeline: acquire a corpus, pretrain, finetune, instruction-tune, align, optimise and then infer §1.3.

Method

This is a literature survey. It has no algorithm of its own. Its structure is the method, so an implementer needs its definition, its categories and its feedback loop.

1. Definition. An agentic LLM receives natural-language input from its environment, reasons to make a decision, and takes autonomous actions that change the environment toward a goal §1.1. The paper says such an agent may plan, reflect on its actions, use tools, keep a model of the world, have internal multi-agent structure, assist humans, live in a society, and create its own training data §1.1.
2. Selection. A Google Scholar search on "Agentic LLM", then snowballing through citations. Only LLM-based work is included, plus some non-agentic LLM papers as background. Most works are from 2024, some from 2023 and 2025. Unrefereed preprints are included, filtered for reputable labs §1.2.
3. Reasoning §2 tab. 1. Works that improve how an LLM decides.
   • 2.1 Multi-step reasoning: chain of thought, search trees, and interpreter or debugger loops §2.1.
   • 2.2 Self-reflection: prompt improvement and reflection by the LLM itself §2.2.
   • 2.3 Retrieval augmentation: fetching outside text at inference time §2.3.
   • 2.4 Discussion: the authors frame these as neurosymbolic, joining neural networks with symbolic reasoning, planning and retrieval at inference time §2.4.
4. Acting §3 tab. 2. Works that connect the model to the world.
   • 3.1 Action models: world models, and vision-language-action (VLA) models trained on robot action sequences §3.1.
   • 3.2 Robots and tools: robot planning, tool calling through APIs, computer and browser tools, and safety and security §3.2.
   • 3.3 Assistants for shopping, flight operations, medicine, science and trading §3.3.
   • 3.4 Discussion, including an in-depth look at AI-scientist systems §3.4.
5. Interacting §4 tab. 3. Works where several agents or an LLM society interact.
   • 4.1 Social capabilities of non-agentic LLMs: conversation, strategic behaviour, social dilemmas and theory of mind §4.1.
   • 4.2 Role-based interaction: teams of LLMs given roles to solve a task together §4.2.
   • 4.3 Open-ended societies: simulacra, emergent social norms and open-world agents §4.3.
   • 4.4 Discussion: scale, emergent collective behaviour and new training data §4.4.
6. Feedback loop. The authors describe the three categories as a cycle. Reasoning enables acting, acting enables interaction, and the outcomes of acting and of social simulations can augment reasoning. They can also be used to finetune LLMs, provided they are validated or filtered §1.1 §5.1 fig. 1. The loop is written as pseudocode below.

# Feedback loop of an agentic LLM, as the paper describes it  [§1.1] [§5.1]
# Input:  model (LLM used as the policy), environment, goal, steps
# Output: validated trajectories for later training or inference-time use

def agentic_loop(model, environment, goal, steps):
    # Decide the next action from the goal and the history so far  [§2]
    trajectory = []
    for step in range(steps):
        action = model.decide(goal, trajectory)
        # Act in the world: tools, robots or other agents respond  [§3] [§4]
        observation = environment.apply(action)
        trajectory.append((action, observation))
    # Keep only grounded results, checked by a human or automatically  [§1.1]
    validated = filter_grounded(trajectory)
    # Reuse for inference-time context or for finetuning  [§5.1]
    return validated

The paper's only formal statement is an informal comparison in §5.1: a self-supervised model predicts a label from an input, while in reinforcement learning a policy picks actions that change the state of an environment. The authors propose that an LLM can serve as that policy.

Hyperparameters

Not applicable. The paper is a literature review that selects and summarises other work §1.2, so it trains and evaluates nothing of its own and states no learning rates, sizes, budgets or settings.

Results

No results of the authors' own. The survey counts its table rows as its coverage tab. 1 tab. 2 tab. 3:

Total: 102 rows. Each row names an approach, its reasoning technique, its application (acting) and its interaction setting. Most rows point to a benchmark, a task or a system rather than to a measured result, and the paper gives no scores for them tab. 1 tab. 2 tab. 3.

The research agenda has four topics, each with concrete directions §5.1 tab. 4:

• Training data: finetune on inference-time reasoning data, make reinforcement learning stable, and add multimodal signals such as speech to VLA training.
• Hallucination: use self-verification, mechanistic interpretability and open-ended or open-world models for exploration.
• Agent behaviour: scalable simulation infrastructure, distilling reasoning into small models, and models of human behaviour at scale.
• Self-reflection: in-model reflection and metareasoning, metacognition and personality, and automated scientific discovery.
• Safety: responsibility and liability for assistants, privacy and fairness of data, and wider application areas.

Paper vs code

No official code. The paper's abstract and the arXiv metadata this card was built from name no code repository. The only project link in the TeX source is a survey website, and it is in a commented-out line that is not in the compiled paper agenticllm4.tex. A survey has no implementation to compare, so this card skips the GitHub search by convention for survey papers.

Pitfalls

• Categories overlap. Most approaches focus on one aspect, and each row shows its main category §1.5 tab. 1 tab. 2 tab. 3. Work that spans categories is placed by its main focus, so look in the other tables too.
• Mixed table entries. The columns mix methods, benchmarks, products and application areas. Many Interacting cells just say "Benchmark" tab. 1 tab. 3. Read each row as a pointer to a primary paper.
• Coverage is not systematic. One search phrase and snowballing, LLM-based work only, unrefereed preprints included §1.2. Expect omissions.
• Author summaries, not checked here. Claims such as zero-shot generalisation of VLAs in domestic tasks §5.1 are the authors' summaries of cited work. Check the cited paper before relying on them.
• The agenda is a list of directions. It names methods such as self-verification and mechanistic interpretability but defines no protocol or metric for them §5.1.
• Safety is open. The authors call it far from solved and give only general directions, such as multi-layer safeguards and standard benchmarks like AgentHarm §5.1.
• Version. The source fetched corresponds to arXiv v3 (updated 2025-11-22 in the metadata). Later versions may change the tables.
• Numbering. Section, table and figure numbers follow the TeX source order. No compiled PDF was checked, so match citations by section title if the rendered numbers differ.
• No numbered equations. Only the inline comparison in §5.1 is formal. Cite it as §5.1, not as an equation number.
• No figures reproduced. Figure content is described only by reference; fig. 1 is the feedback cycle and fig. 3 the taxonomy.

Lineage

• Builds on: older agent research in symbolic AI, game theory, robotics, multi-agent systems and reinforcement learning, which the paper uses to define agency §1.1.
• Compared against: related surveys the paper names, each covering part of the field: Li et al. (2024) on retrieval and tool use, Wang et al. (2024) on autonomy and agent construction, Gao et al. (2024) on multi-agent modelling and simulation, Xi et al. (2023) on interactive agents, and Zhao et al. (2023) as a general LLM survey §1.2.
• Superseded by: not stated.