# Architectures for Building Agentic AI

**Authors:** Sławomir Nowaczyk  
**arXiv:** [2512.09458](https://arxiv.org/abs/2512.09458)  
**Date:** 2025-12-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2512.09458v1/

## Key points

- Fundamental architectural aspects contributing to reliability of AI systems:
- Key components: goal manager, planner, tool-router, executor, memory, verifiers, safety monitor, telemetry
- Importance of structured interfaces with schema constraints and validation mechanisms
- Tool-using agents
- Memory-augmented agents
- Planning and self-improvement agents
- Multi-agent systems
- Embodied/web agents
- Typed schemas
- Idempotency
- Permissioning
- Transactional semantics
- Memory hygiene
- Runtime governance
- Simulate-before-actuate safeguards
- Discussion on orchestration frameworks like LangGraph and AutoGen:
- Enforcing contracts through typed messages
- Implementing budget constraints per edge or role capabilities
- Architecture plays a critical role in determining the evolution of a dependable system. Models propose while architectures dispose.

## Summary

This chapter focuses on the fundamental architectural aspects that contribute to the reliability of AI systems. The discussion revolves around key components such as the goal manager, planner, tool-router, executor, memory, verifiers, safety monitor and telemetry. It emphasizes the importance of structured interfaces with schema constraints and validation mechanisms to reduce silent failures in downstream processes. A taxonomy of AI patterns is introduced including tool-using agents, memory-augmented agents, planning and self-improvement agents, multi-agent systems and embodied/web agents. Each pattern is analyzed in terms of its impact on reliability and potential failure modes. Practical design guidance is provided on typed schemas, idempotency, permissioning, transactional semantics, memory hygiene, runtime governance and simulate-before-actuate safeguards. Furthermore,the chapter discusses orchestration frameworks such as LangGraph for stateful graph runtimes and AutoGen for conversation-based computation in single- and multi-agent settings. It highlights the importance of enforcing contracts through typed messages and implementing budget constraints per edge or role capabilities to ensure controlled evolution over time. The key takeaway message from this chapter is that architecture plays a critical role in determining whether a powerful reasoning core can evolve into a dependable system. By emphasizing isolation through components, contracts through schemas and validators,and controlled evolution through orchestration frameworks; it is emphasized that models propose while architectures dispose. This sets the stage for subsequent chapters to delve into specific topics like data coverage estimation,cybersecurity monitoring,and governance within the framework of well-designed interfaces established earlier.