AWS re:Invent 2025 - Agentic AI Meets responsible AI: Strategy and best practices (AIM422)

Summary of "Agentic AI Meets Responsible AI: Strategy and Best Practices"

Introduction to Agentic AI

• Agentic AI refers to AI systems with varying degrees of autonomy, from simple turn-taking interactions to agents that can plan and execute complex tasks independently.
• Examples of agentic AI include:
  • Agentic executive assistants that can manage schedules, book travel, and make purchases on behalf of users.
  • Planning agents that can take high-level goals and autonomously plan and execute the necessary steps.

Scientific Challenges of Agentic AI

Embeddings and the Native Language of AI

• AI models, including agentic systems, operate on data representations called embeddings rather than the raw input data.
• Embeddings map data into a lower-dimensional space that preserves similarities, but the semantics of this space are not well-understood.
• This can have implications for privacy and security, as the boundaries between different types of data (e.g., financial, medical) may become blurred in the embedding space.
• There is limited research on how to reverse-engineer sensitive information from embeddings, posing challenges for privacy-preserving design.

Negotiation and Collective Behavior

• Experiments have shown that large language models exhibit similar negotiation and bargaining behaviors to humans in games like the Ultimatum Game.
• This suggests agentic AI systems may exhibit emergent collective behaviors, like collusion or systemic risk, that are difficult to predict from individual agent actions.
• Understanding and mitigating these collective effects will be crucial as more economic and strategic activities are mediated by agentic AI.

Subjective Common Sense

• Humans have a wealth of "subjective common sense" heuristics and preferences that are personal and difficult to codify, such as habits around locking doors or sharing passwords.
• Imbuing agentic AI with this type of personalized common sense knowledge will be a significant challenge, as it cannot be easily learned from data like more objective common sense facts.

Evolving Agentic AI Capabilities

Trends in Agentic AI Development

1. Fast to Slow: Agents will transition from handling quick, transactional tasks to longer-term, continuous planning and execution.
2. Reactive to Proactive: Agents will infer and initiate tasks without explicit user prompting, based on contextual awareness.
3. Single Vendor to Multi-Party Connectivity: Agents will need to seamlessly integrate with third-party workflows and systems, posing new challenges.

Responsible AI Considerations

• Agentic AI systems require a shift in mindset from traditional software, as they are based on statistical models rather than deterministic logic.
• Key differences include:
  • Performance is data-dependent, not guaranteed across all inputs.
  • Multiple models can achieve similar performance, with different internal "logic".
  • Developers must make explicit decisions about technical properties like privacy, security, and fairness.
• Building agentic AI systems requires a framework for benefit-risk analysis, release criteria definition, data set planning, and system architecture design to address these challenges.

Conclusion

As agentic AI capabilities continue to evolve, responsible development practices will be crucial to realize the benefits while mitigating the risks. Key focus areas include understanding the unique properties of AI systems, designing for privacy and security, and addressing the challenges of collective agent behaviors and personalized common sense.