Session brief
Cohere: Building enterprise LLM agents that work (Shaan Desai)
Overview
This talk focuses on the practical challenges and solutions for building enterprise-grade LLM agents. It emphasizes that while agents are a powerful application of generative AI, creating them at scale, safely, and seamlessly is complex due to the vast array of frameworks, tools, and evaluation criteria. The presentation aims to guide builders through critical decision-making processes, sharing key learnings from Cohere's experience in developing these agents.
Who should watch
- AI Engineers
- Product Managers
- Builders working with LLM agents
- Those facing challenges in scaling, safety, and seamless integration of LLM agents
- Individuals looking for insights into framework selection, agent strategies, and evaluation methods
Key takeaways
- Framework selection for enterprise agents should prioritize observability, setup cost (iteration speed), and support (documentation, model/tool compatibility). Native solutions or LangChain are recommended for high observability, while frameworks like AI and AutoGen are suitable for quick proofs-of-concept due to low setup costs and ease of orchestrating pre-built agents.
- When choosing an agent strategy, start simple with a single LLM and a few tools. Diligent attention to tool specifications, including clear descriptions, simplified input types, and concise instruction lists, significantly improves performance.
- Long chat histories (over 20 turns) can induce hallucinations; caching and retrieving relevant history based on new queries can improve agent performance.
- For multi-agent systems, the router should have clear tool descriptions and sharp routing instructions to handle edge cases. Sub-agents should be constrained to performing independent tasks with a small set of tools.
- Human-in-the-loop is critical for business applications, allowing for codified rules to trigger user confirmation before or after tool calls, especially when execution outputs might be sensitive.
- Evaluating agent performance requires assessing intermediate stages, not just the final output. Building a golden set of ground truth queries, expected function calls, parameters, and responses is crucial for identifying and debugging failure points.
- Failure mitigation strategies vary by severity: prompt engineering for low-severity failures, targeted annotation datasets for 10-20% failures, and synthetic data generation with fine-tuning for high failure rates or complex disambiguation tasks.
- Cohere has developed a product called North, a single-container deployment that integrates RAG, vector databases, and connectivity to various applications like Gmail, Outlook, Drive, and Slack, serving as a comprehensive platform for building and using agentic applications.
Notable quotes
*Building large scale Enterprise agents often requires high levels of observability for which we would really recommend going native or building with Langra.*
*Our core recommendation and this is insights that have come from a number of use cases is always start simple a single LLM with a handful of tools can often go a long way.*
*Autonomous LLM agents do indeed have a tendency to fail.*
Unofficial community note. Prefer the recording for nuance.