Lessons Learned from Building Agentic Systems With Jayeeta Putatunda

Krishna Gade hosts a discussion with Jayeeta Putatunda, focusing on the challenges and insights related to building agentic systems. Jayeeta highlights the difficulty in measuring improvements in AI systems and emphasizes the 80-20 rule, advocating for prioritizing impactful use cases over extensive prototyping. She notes the rapid obsolescence of AI frameworks and models, complicating the evaluation of generative AI systems.

Jayeeta defines agentic solutions through use cases and available frameworks, stressing the importance of identifying high ROI patterns. She expresses skepticism about the industry's readiness for high autonomy use cases in finance, underscoring the need for credibility and accountability in the autonomy pipeline. The discussion explores two types of agentic systems: deterministic workflow-like systems and fully autonomous agents that incorporate reflection. Jayeeta explains that while the first type is manually constructed, the second automates workflows and self-corrects, guided by business rules.

Key considerations include the necessity of checkpoints in workflows, the importance of reflection for large language models (LLMs), and the potential biases that may arise from granting too much autonomy to reflection agents. Jayeeta emphasizes responsible system design that minimizes bias and adheres to business guidelines, suggesting third-party evaluations for unbiased assessments. The challenges of testing agentic systems due to their non-deterministic nature are acknowledged, along with the critical role of governance, observability, and data preparation in effective system evaluation.

The speaker emphasizes the need to break down evaluation metrics into multiple components, as there is no single metric that can effectively track performance. Sufficient logging metrics are crucial for assessing the correctness and quality of links suggested by deep research agents. The complexity of using multiple models and tools in an agenting system necessitates tracking token usage, response times, and error rates to ensure a positive user experience.

To diagnose failures effectively, implementing checkpoints at each step of an agent's process is recommended. This includes logging inputs, outputs, and responses extensively, as even seemingly excessive logging can yield valuable insights. Observability in agentic AI should begin with defining metrics and logging from the outset. Human oversight is vital for identifying patterns in outputs and enhancing system scalability.

The extraction process faces challenges due to the complexity of financial documents, which often combine text, infographics, and tables. A deterministic approach is necessary in financial contexts, where inaccuracies in numerical data are unacceptable. Addressing the non-deterministic outputs from LLMs is critical, as these models may struggle with quantitative understanding. There is a call for integrating classical machine learning models with LLMs to create hybrid models that leverage past learnings.

An audience member raises concerns about the reliability of agentic systems in production, noting that while systems may perform well in test cases, they struggle with noisy inputs in real-world environments. Thorough testing during QA and beta phases is essential to identify potential issues before production release. Engaging beta testers to push applications to their limits can refine products and garner stakeholder support.

Agentic systems generate extensive logs that include reasoning, API calls, and planning steps, allowing for effective root cause analysis. Building agentic systems requires an iterative approach, starting with three to four basic components and continuously refining evaluation criteria. Collaboration is vital in the development process, particularly for customer support agents, ensuring that logging of all components is in place.

The evolution of operations in this field has transitioned from MLops to LLMops to AgentOps, with foundational concepts like system usefulness, accuracy, and uptime remaining consistent. Key metrics must ensure that the system is relevant and accurately responds to user queries. Challenges in AgentOps include system design and agent coordination, emphasizing the need for careful orchestration in multi-agent systems.

For new AI teams deploying their first agentic applications, it is important to identify the expected output and the user problem being solved. Assessing existing systems for data and process gaps, utilizing open-source tools, and establishing checkpoints with subject matter experts (SMEs) are key steps. Certain use cases may not be suitable for agents, particularly low-risk scenarios with minimal output variance.

To prevent agents from failing in front of clients and to build trust with stakeholders, it is essential to educate them about the unpredictability of AI systems. Building trust also involves emphasizing evaluation, observability, and guardrails in non-deterministic software. Highlighting productivity improvements underscores the need for modular structures in software development, advocating for simplicity in design and applying the 80/20 rule to prioritize impactful use cases.