LeiWangetal. ASurveyonLargeLanguageModelbasedAutonomousAgents 25
Table2 RepresentativeapplicationsofLLM-basedau- seekclarification,andincorporatehumanfeedback,
tonomousagents.
showing the potential for human-AI collaboration
Domain Work
inengineeringdesign.
TE[102],Akataetal.[103],Ziems
Psychology
etal.[105],Maetal.[104]
Computer Science & Software Engineering:
Out of One [29], Horton [106],
Political
Ziemsetal.[105] Inthefieldofcomputerscienceandsoftwareengi-
Scienceand
Social
neering, LLM-based agents offer potential for au-
Economy
Science
Social Simulacra [79], Gen-
tomating coding, testing, debugging, anddocumen-
erative Agents [20], SocialAI
tation generation [14,18,23,24,126–128]. Chat-
Social School [109], AgentSims [34],
Simulation S3 [77],Williamsetal.[110],Li Dev[18]proposesanend-to-endframework,where
etal.[107],Chaoetal.[108]
multiple agent roles communicateand collaborate
ChatLaw [112], Blind Judge-
Jurisprudence
ment[113] through natural language conversations to complete
Research Ziemsetal[105],Bailetal.[114]
the software development life cycle. This frame-
Assistant
workdemonstratesefficientandcost-effectivegener-
Documentation ChemCrow[75],Boikoetal.[115]
andData ationofexecutablesoftwaresystems. ToolBench[14]
Management
canbeusedfortaskssuchascodeauto-completion
ChemCrow[75],Boikoetal.[115],
Experiment
Natural Grossmannetal.[122] andcoderecommendation. MetaGPT[23]abstracts
Assistant
Science
ChemCrow[75],CodeHelp[120], multipleroles,suchasproductmanagers,architects,
Natural Boiko et al. [115], MathA- projectmanagers,andengineers,tosupervisecode
Science gent[117],Drorietal.[118]
generation process and enhance the quality of the
Education
RestGPT [70], Self- final output code. This enables low-cost software
collaboration [24], SQL-
development. [24] presents a self-collaboration
PALM [90], RAH [92], DB-
framework for code generation using LLMs. In
GPT [41], RecMind [51],
CS&SE ChatEDA [123], InteRecA- thisframework,multipleLLMsareassumedtobe
gent [124], PentestGPT [125],
distinct "experts" for specific sub-tasks. They col-
Engineering CodeHelp [120], SmolMod-
laborateandinteractaccordingtospecifiedinstruc-
els [126], DemoGPT [127],
GPTEngineer[128] tions, forming a virtual team that facilitates each
GPT4IA [129], IELLM [130],
Industrial other’swork. Ultimately,thevirtualteamcollabo-
TaskMatrix.AI[71]
Automation
ratively addresses code generation tasks without re-
ProAgent [131], LLM4RL [132],
PET[133],REMEMBERER[134], quiringhumanintervention. LLIFT[139]employs
Robotics& DEPS [33], Unified Agent [135], LLMsto assistinconductingstaticanalysis, specif-
Embodied SayCan[78],LMMWM[136],Tidy-
icallyforidentifyingpotentialcodevulnerabilities.
AI Bot[137],RoCo[93],SayPlan[31]
This approacheffectivelymanages the trade-offbe-
complexstructuressuchasbuildings,bridges,dams, tweenaccuracyandscalability. ChatEDA[123]is
roads,etc.[138]proposesaninteractiveframework anagentdevelopedforelectronicdesignautomation
where human architects and agents collaborate to (EDA)tostreamline thedesignprocess byintegrat-
constructstructuresina3Dsimulationenvironment. ingtask planning,scriptgeneration, andexecution.
The interactive agent can understand natural lan- CodeHelp [120] is an agent designed to assist stu-
guage instructions, place blocks, detectconfusion, dentsanddevelopersindebuggingandtestingtheir