AWS re:Invent 2025 - Function calling vs agents: Choose the right AI approach (DEV204)

Summary of AWS re:Invent 2025 - Function calling vs agents: Choose the right AI approach (DEV204)

Introduction

• The presentation discusses the design choices companies face when integrating generative AI into real-world systems.
• Two main approaches are explored: function calling within large language models (LLMs) and deploying autonomous AI agents.
• Each approach offers different benefits and trade-offs in terms of complexity, performance, safety, and business value.

Function Calling

• Function calling allows developers to define APIs and functions that an LLM can invoke to access external data, perform calculations, or interact with code.
• This provides more reliable and deterministic outputs compared to unconstrained text generation, making AI safer to integrate into business workflows.
• Example use case: A banking assistant can retrieve a customer's exact account balance by calling a backend function, rather than guessing the answer.
• Function calling is efficient for simple, known interactions, but becomes more complex as the number of interdependent functions and dynamic requirements scale.

AI Agents

• Agents are autonomous AI processes that can reason about how to solve a problem, decompose tasks into multiple steps, and dynamically choose which tools or functions to use.
• Agents are more adaptable than function calling, as they can plan multiple steps ahead, analyze intermediate results, and adjust their actions accordingly.
• This allows agents to actively drive problem-solving, rather than just passively responding.
• Example use case: A travel planning agent can autonomously retrieve weather forecasts, check flight and hotel availability, and suggest activities based on the user's needs and the context.

Multi-Agent Systems

• For more complex problems that span multiple domains, a single agent may not be enough, and a multi-agent system can be more effective.
• In a multi-agent system, specialized agents focus on particular areas, with their own expertise, tools, and instructions, and a coordinator agent orchestrates their collaboration.
• Two common multi-agent architectures are the hierarchical (supervisor-worker) pattern and the network (swarm) pattern, each with their own trade-offs.

Challenges and Considerations

• Key challenges with multi-agent systems include orchestration complexity, latency and costs, debugging, and security/safety concerns.
• When choosing between function calling, single agents, and multi-agent systems, the main factors to consider are:
  • Task complexity: Simple tasks may only require function calling, while more complex, multi-step problems may benefit from agents or multi-agent systems.
  • Predictability vs. adaptability: Function calling offers more predictable behavior, while agents and multi-agent systems allow for more adaptability and emergent problem-solving.
  • Cost and latency: Function calling is generally faster and cheaper, while agents and multi-agent systems can be more resource-intensive.
  • Control over business logic: Function calling keeps the logic fully under the developer's control, while agents and multi-agent systems delegate more decision-making to the AI.
  • Autonomy: Function calling is reactive, agents are proactive, and multi-agent systems enable collaborative, autonomous problem-solving.

Real-World Use Cases

• Function calling: FAQ bots, currency conversion, order status lookups
• Single agents: Travel planning, personal research assistants, content summarization
• Multi-agent systems: Complex contract generation, enterprise-wide troubleshooting, legal due diligence

Key Takeaways

1. Start simple with function calling for deterministic, straightforward tasks.
2. Use single agents for multi-step tasks that require adaptive reasoning.
3. Employ multi-agent systems for complex problems involving multiple domains and specialized collaboration.
4. Always plan for observability, monitoring, and governance when deploying AI-powered systems.
5. Leverage AWS services like Amazon Bedrock, Lambda, Sagemaker, and Step Functions to build end-to-end AI solutions.