Abstract
Foundation model-enabled generative artificial intelligence facilitates the development and implementation of agents, which can leverage distinguished reasoning and language processing capabilities to take a proactive, autonomous role in pursuing users' goals. Nevertheless, there is a lack of systematic knowledge to guide practitioners in designing agents considering challenges of goal-seeking, such as hallucinations inherent in foundation models, explainability of reasoning processes, and complex accountability. To address this issue, we performed a systematic literature review to understand the state-of-the-art foundation model-based agents and the broader ecosystem. In this paper, we present a pattern catalogue consisting of 18 architectural patterns with analyses of the context, forces, and trade-offs. The proposed catalogue provides holistic guidance for the effective use of patterns and supports the architecture design of foundation model-based agents by facilitating goal-seeking and plan generation.
Introduction
While autonomous agents like AutoGPT and BabyAGI have sparked immense interest, building them for production remains a challenge. Developers face critical hurdles including reasoning uncertainty (hallucinations), underspecification of user goals, and the "black box" problem, which limits transparency and trust. This paper addresses the need for a systematic "language" or framework to guide the architectural design of these systems, moving beyond ad-hoc prompting toward structured software engineering.
Methodology
The researchers conducted a Systematic Literature Review (SLR), analyzing 57 high-quality academic studies to extract recurring solutions to common agent design problems. These findings were synthesized into a formal catalogue using a standardized pattern template covering Context, Problem, Forces, Solution, and Consequences. The study also incorporated "grey literature" from industry-leading projects like LangChain and OpenAI to identify real-world "known uses."
The Agent Design Pattern Catalogue
The core contribution is the categorization of 18 architectural patterns into functional groups that define how agents interact with users, models, and each other. These groups include Goal Creation, Querying & Planning, Reflection, Cooperation, and Guardrails & Infrastructure.
Key Insights and Trade-offs
- The Proactive Shift: There is a notable transition from Passive Goal Creators (reacting to prompts) to Proactive Goal Creators that use environmental detectors to anticipate user needs, a key step toward ambient intelligence.
- Querying Trade-offs: The catalogue formalizes the choice between One-shot Model Querying (cost-efficient and simple) and Incremental Model Querying (higher reasoning certainty and explainability, but slower and more expensive).
- The Reflection Hierarchy: Reliability is improved through a hierarchy of feedback: Self-reflection (internal), Cross-reflection (agent-to-agent), and Human-reflection (incorporating user feedback).
- Multi-Agent Governance: Structured cooperation patterns like Debate-based or Role-based Cooperation facilitate division of labor and improve critical thinking by reaching consensus through adversarial or specialized interactions.
Conclusion
The primary outcome of this work is a decision model that guides practitioners through the selection process based on project constraints. For instance, limited budgets favor One-shot Querying, while high-stakes environments requiring accuracy and safety necessitate Incremental Querying and Multimodal Guardrails.
Future work will focus on quantifying the performance metrics (latency and token cost) of these patterns and mapping them to international AI regulations, such as the EU AI Act, to ensure compliance by design. Additionally, the researchers plan to explore how these patterns integrate into "Agentic DevOps" workflows for continuous evaluation.
Summary updated