Session brief
Scaling Agents for Gen AI Products - Anju Kambadur, Bloomberg Head of AI Engineering
Overview
This talk focuses on the practical challenges and strategies for scaling generative AI agents in product development, particularly within a large financial data organization. It emphasizes the need for robust engineering practices, acknowledging the inherent fragility and evolving nature of LLMs and agentic systems. The core thesis is that building reliable AI products requires moving beyond basic LLM capabilities to implement structured development, resilient architectures, and clear organizational design.
Who should watch
- AI Engineers
- Product Managers
- Builders working with LLMs and agentic systems
- Those facing challenges in productizing AI
- Individuals interested in organizational structures for AI development
Key takeaways
- Building AI products requires a shift from internal LLM development to leveraging and building upon existing open-source and open-weight models.
- The inherent stochasticity and error multiplication in LLM compositions lead to fragile behavior, necessitating built-in guardrails and safety checks rather than relying solely on upstream system accuracy.
- Scaling agents requires a mindset shift towards resilience, allowing individual agents to evolve without constant handshakes with all downstream consumers.
- Organizational structure is crucial; initially, collapsing the stack for fast iteration is beneficial, but as understanding grows, moving towards horizontally aligned functions (like guardrails) and breaking down monolithic agents into smaller, optimized pieces becomes necessary.
- Traditional ML product development, with well-defined APIs and predictable inputs/outputs, contrasts sharply with the LLM landscape, where robust monitoring, remediation workflows, and circuit breakers are essential.
- Non-negotiable principles like precision, comprehensiveness, speed, and data protection must ground any AI product development, regardless of the technology used.
- A semi-agentic approach, where certain autonomous pieces are balanced with non-negotiable, coded-in guardrails, is currently a practical strategy for building trust and ensuring factual accuracy.
- The evolution of data infrastructure, from sparse to dense and hybrid indices, supports the development of more sophisticated agentic architectures.
Notable quotes
*The errors multiply a lot and that is something that causes a lot of fragile behavior.*
*It is easier to not count on the Upstream systems to be accurate but rather factor in that they will be fragile and they'll be evolving and just do your own safety checks.*
*We don't want every team every one of those 50 teams like trying to figure out what does it mean for me to not accept user inputs that are thinly wailed Financial advice inputs.*
Unofficial community note. Prefer the recording for nuance.