AI Engineering: Building Better Platforms

Juan Herreros Elordoy, a team lead for cloud-native technology at Banking Circle, recently shared insights into the development and deployment of AI agents, emphasizing the importance of robust platform engineering. In his presentation, titled "Platforms for Humans and Machines," Elordoy detailed the key principles and practices that contribute to successful AI agent integration within an organization.

Banking Circle's Platform Engineering Approach

Banking Circle, a provider of global cross-border payments, accounts, and liquidity management, processes over 1 trillion euros annually. The company's commitment to technology is reflected in its workforce, with 40% of its employees being engineers. Elordoy highlighted that Banking Circle established a platform engineering team several years ago to manage its technological infrastructure.

The "Atlas" Platform

This platform engineering team, codenamed "Atlas," comprises over 250 individuals dedicated to building critical systems. Their work includes developing client-facing APIs, core banking systems, internal tools, data infrastructure, and integrations with other companies and clearing schemes. Elordoy stressed that these teams are focused on building systems that are both technically sound and user-friendly.

Key Principles for AI Agent Platforms

Elordoy outlined six key principles for building effective platforms that can support AI agents:

• Self-service: Platforms should be automatic, intuitive, and require no human involvement for users to perform tasks.
• API-based: The platforms must be discoverable, with schema validation and authentication, allowing agents to interact programmatically.
• Local-first: Developers should be able to shift left, meaning they can test and deploy agents locally until they are confident in their functionality.
• Documentation: Comprehensive documentation, including context, code examples, and API details, is crucial for agent usability.
• Enable Contributions: Platforms should encourage contributions by providing guardrails for security, compliance, and standards, and offering context for ease of use and developer experience.
• Measure: It's essential to measure the effectiveness of AI agents using metrics like delivery (DORA), reliability (SLOs, golden signals), support requests, and developer experience (SPACE).

The Human vs. AI Agent Experience

Elordoy illustrated the difference between a successful and a struggling developer's experience when interacting with AI agents. He emphasized that while developers can write code, deploying and verifying it can be challenging without the right platform. AI agents, he suggested, can assist in this process by automating tasks and providing clear feedback loops.

He drew an analogy to a "Trojan Horse" for AI, suggesting that while AI can bring many benefits, it's crucial to have best practices in place to manage its integration effectively. The analogy implies that AI can be a powerful tool, but its implementation needs careful consideration to avoid unintended consequences.

Measuring Success and Fostering Collaboration

Elordoy highlighted the importance of measuring the impact of AI agents, not just in terms of technical performance but also in developer experience. He suggested that by providing clear documentation and enabling self-service through APIs, organizations can reduce the friction for developers and accelerate the adoption of AI.

He also touched upon the concept of "shifting left" in development, encouraging developers to test and validate their work as early as possible in the process. This, he argued, leads to more robust and reliable systems.

The Role of Documentation and Metrics

The presentation underscored that clear documentation is paramount for AI agents to be effective. Agents need to understand their tasks, the available resources, and how to verify their results. By providing this context, developers can ensure that AI agents are used efficiently and securely.

Furthermore, Elordoy stressed the importance of measuring the success of these platforms and agents. Metrics related to delivery, reliability, support requests, and developer experience provide valuable insights into what is working and what needs improvement.

Conclusion

Elordoy's presentation provided a comprehensive overview of building and managing AI agent platforms. The key takeaway is that while AI agents offer significant potential, their successful integration requires a strategic approach focused on self-service, API accessibility, thorough documentation, robust metrics, and fostering a culture of collaboration and continuous improvement.