Elamur Research⌕
☰
home / papers / 2024 / The Landscape of Emerging AI Agent Architectures for Reasoning, Planning, and Tool Calling: A Survey

The Landscape of Emerging AI Agent Architectures for Reasoning, Planning, and Tool Calling: A Survey

Paper · 2024 · Tula Masterman, Sandi Besen, Mason Sawtell et al.

Summary

This survey by Masterman, Besen, Sawtell and Chao (2024) reviews how LLM-powered agents are built to reach complex goals that require reasoning, planning and calling external tools. It does not propose a new system. Instead, it compares existing designs. Single-agent setups rely on one model for all work. Multi-agent setups sit on a spectrum from vertical, where a leader coordinates subordinates, to horizontal, where peers collaborate in one shared conversation. The authors examine representative methods, including ReAct, RAISE, Reflexion, AutoGPT + P and LATS. They discuss how leadership, communication style and separate planning, execution and reflection stages shape results. Their main finding is that no single architecture is best for every task, and the right choice depends on the use case. Strong systems tend to share certain traits: clear system prompts, explicit division of work, distinct reasoning and evaluation phases, human or agent feedback, and filtering of messages. The paper also highlights open problems in evaluation, such as benchmark contamination, weak transfer to real-world data, and bias. This makes it a practical orientation map for teams designing agent pipelines.

Key points

  • The survey groups planning approaches into five families: breaking a task into smaller parts, choosing among several candidate plans, planning with the help of external modules, reflecting on and refining plans, and planning supported by memory.
  • On HotpotQA, ReAct produced hallucinations in about 6% of cases, compared with 14% for chain-of-thought (CoT). However, it can get stuck repeating the same thoughts and actions.
  • RAISE extends ReAct with memory. It uses a scratchpad as short-term memory and a store of similar past examples as long-term memory, which improves context retention in long dialogues.
  • Single-shot benchmarks such as MMLU and GSM8K are poor proxies for agent ability. Agent-oriented alternatives include AgentBench, SmartPlay, StrategyQA, SWE-bench (real GitHub issues in Python) and WildBench, which draws on roughly 570,000 real ChatGPT conversations from WildChat.
  • Agents can show more harmful behaviour and stronger social bias than plain LLMs. The authors argue that robust bias evaluation must involve human judges rather than only LLM-as-judge methods.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 19,503 worksArtificial Intelligence · 18,805 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,948 works

Cite

Cite as Tula Masterman, Sandi Besen, Mason Sawtell and Alex Chao. 2024. The Landscape of Emerging AI Agent Architectures for Reasoning, Planning, and Tool Calling: A Survey. arXiv:2404.11584.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2404.11584 · arXiv 2404.11584

@misc{masterman2024landscape,
  title = {{The Landscape of Emerging AI Agent Architectures for Reasoning, Planning, and Tool Calling: A Survey}},
  author = {Tula Masterman, Sandi Besen, Mason Sawtell and Alex Chao},
  year = {2024},
  doi = {10.48550/arxiv.2404.11584},
  eprint = {2404.11584},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2404.11584}
}

References

12 references, 7 of them with a card of ours. ↗ leads to the work.

  1. Timo Birr et al. AutoGPT+P: Affordance-based Task Planning with Large Language Models. arXiv:2402.10778 [cs] version: 1. Feb. 2024. URL : http://arxiv.org/abs/2402.10778. Weize Chen et al. AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors. arXiv:2308.10848 [cs]. Oct… ↗
  2. Zhenhailong Wang et al. Unleashing the Emergent Cognitive Synergy in Large Language Models: A Task-Solving Agent through Multi-Persona Self-Collaboration. 2024. arXiv: 2307.05300 [cs.AI].
  3. Jason Wei et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv:2201.11903 [cs]. Jan. 2023. URL : http://arxiv.org/abs/2201.11903. ↗
  4. Yue Wu et al. SmartPlay: A Benchmark for LLMs as Intelligent Agents. arXiv:2310.01557 [cs]. Mar. 2024. URL : http://arxiv.org/abs/2310.01557. ↗
  5. Zhiheng Xi et al. The Rise and Potential of Large Language Model Based Agents: A Survey. 2023. arXiv: 2309.07864 [cs.AI].
  6. Shunyu Yao et al. ReAct: Synergizing Reasoning and Acting in Language Models. arXiv:2210.03629 [cs]. Mar. 2023. URL : http://arxiv.org/abs/2210.03629. ↗
  7. Shunyu Yao et al. Tree of Thoughts: Deliberate Problem Solving with Large Language Models. arXiv:2305.10601 [cs]. Dec. 2023. URL : http://arxiv.org/abs/2305.10601. ↗
  8. Muru Zhang et al. How Language Model Hallucinations Can Snowball. arXiv:2305.13534 [cs]. May 2023. URL : http://arxiv.org/abs/2305.13534. ↗
  9. Wenting Zhao et al. “(InThe)WildChat: 570K ChatGPT Interaction Logs In The Wild”. In: The Twelfth International Conference on Learning Representations. 2024. URL : https://openreview.net/forum?id= Bl8u7ZRlbM. ↗
  10. Andy Zhou et al. Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models. arXiv:2310.04406 [cs]. Dec. 2023. URL : http://arxiv.org/abs/2310.04406. ↗
  11. Kaijie Zhu et al. DyVal 2: Dynamic Evaluation of Large Language Models by Meta Probing Agents. arXiv:2402.14865 [cs]. Feb. 2024. URL : http://arxiv.org/abs/2402.14865. ↗
  12. Kaijie Zhu et al. DyVal: Dynamic Evaluation of Large Language Models for Reasoning Tasks. arXiv:2309.17167 [cs]. Mar. 2024. URL : http://arxiv.org/abs/2309.17167. ↗

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/masterman-2024-landscape-emerging-ai-agent-architectures/