Digitala Vetenskapliga Arkivet

Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Arkitektonisk kontroll i LLM-baserade AI-agenter
Blekinge Institute of Technology, Faculty of Computing.
2026 (Swedish)Independent thesis Basic level (university diploma), 10 credits / 15 HE creditsStudent thesis
Abstract [en]

This study investigates how different agent architectures - single-agent, multi-agent, and human-in-the-loop - affect control, transparency, and reliability in systems based on large language models (LLMs). The study combines a literature review with an experimental prototype in which the three architectures are implemented and evaluated using a set of structured test cases. 

To enable a systematic comparison, a set of measurable evaluation dimensions is defined, including the number of identified issues, the number of implemented improvements, and the presence of factual errors in generated responses. These dimensions operationalize aspects of control and transparency by capturing how each architecture identifies, processes, and corrects deficiencies in its outputs. 

The results indicate that multi-agent architectures enable more structured internal review and iterative refinement, while human-in-the-loop provides the highest level of control and traceability through explicit human intervention. In contrast, the single-agent architecture is simpler but lacks explicit mechanisms for iterative validation and improvement. 

The study also highlights challenges in evaluating AI-generated responses, as the assessment process is partly interpretative and influenced by task characteristics. The findings suggest that there is no universally optimal architecture; instead, the choice involves trade-offs between efficiency, control, transparency, and system complexity.

Abstract [sv]

Denna studie undersöker hur olika agentarkitekturer - single-agent, multi-agent och human-in-the-loop - påverkar kontroll, transparens och tillförlitlighet i system baserade på stora språkmodeller (LLM). Studien kombinerar en litteraturanalys med ett experiment där tre arkitekturer implementeras och utvärderas genom ett antal strukturerade testfall. 

För att möjliggöra en systematisk jämförelse definieras ett antal mätbara analysdimensioner, såsom antal identifierade brister, antal implementerade förbättringar samt förekomst av faktiska fel i genererade svar. Dessa dimensioner används för att operationalisera kontroll och transparens genom att analysera hur olika arkitekturer identifierar, hanterar och korrigerar brister. 

Resultaten visar att multi-agent-arkitekturer möjliggör mer strukturerad intern granskning och iterativ förbättring, medan human-in-the-loop ger högst grad av kontroll och spårbarhet genom explicit mänsklig involvering. Single-agent-arkitekturen uppvisar en enklare struktur men saknar tydliga mekanismer för iterativ validering och förbättring. 

Studien visar även att utvärdering av AI-genererade svar är utmanande, då den delvis bygger på tolkning och påverkas av uppgiftens karaktär. Resultaten indikerar att valet av arkitektur innebär en avvägning mellan effektivitet, kontroll, transparens och systemkomplexitet.

Place, publisher, year, edition, pages
2026. , p. 38
National Category
Computer Sciences
Identifiers
URN: urn:nbn:se:bth-29909OAI: oai:DiVA.org:bth-29909DiVA, id: diva2:2077055
Subject / course
PA1438 Självständigt arbete Webbprogrammering
Educational program
PAGWG Webbprogrammering
Available from: 2026-06-29 Created: 2026-06-22 Last updated: 2026-06-29Bibliographically approved

Open Access in DiVA

fulltext(715 kB)1 downloads
File information
File name FULLTEXT01.pdfFile size 715 kBChecksum SHA-512
687ff6e8f91f20a531d5f31276a0a82d3a6982b03d0ab59c2811015b89e1e5bc35491c43ead474db133d276cab3ea6967e57a1e0f601aaf6f45fed8ea14720d9
Type fulltextMimetype application/pdf

By organisation
Faculty of Computing
Computer Sciences

Search outside of DiVA

GoogleGoogle Scholar
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

urn-nbn

Altmetric score

urn-nbn
Total: 10 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf