Marc Klingen on AI Agents & Langfuse

Marc Klingen, a co-founder of Langfuse, shared insights at AI Engineer London 2023 on the challenges and lessons learned in upskilling AI coding agents. Klingen highlighted the critical need for robust tooling to effectively manage and debug these agents, emphasizing Langfuse's role in providing essential observability.

Companies working on this

StartupHub profiles of the companies this article names, with funding and a one-liner from our database.

Langfuse

$3M

Open-source LLM engineering platform for debugging, monitoring, and improving production-grade LLM applications.

poolside

$4K

AI coding models and pair programmer built for on-device execution with privacy-first architecture.

Sierra

$10K

Sierra helps businesses build better, more human customer experiences with AI agents that can understand, reason, and take action across various channels.

Inflection

$4.0B

Develops enterprise AI solutions and conversational AI models, formerly known for its personal AI assistant, Pi.

The Evolution of Agent Development

Klingen drew a parallel between early attempts to solve the Rubik's Cube and the current state of AI agent development. Initially, agents were like the complex, manual process of solving the cube. However, as tools and methodologies evolved, the process became more streamlined and predictable. Similarly, the development of AI agents is moving from ad-hoc, difficult-to-manage processes towards more structured and observable systems.

Workflows vs. Agents

A key distinction Klingen made was between traditional workflows and agent-based systems. Workflows are characterized by formalized paths to achieve a goal, often involving a series of predefined LLM calls. Agents, on the other hand, offer more freedom, allowing them to find and verify their own way through an environment using feedback loops. This inherent flexibility, while powerful, also necessitates better tools for understanding and controlling their behavior.

The Langfuse Solution

Langfuse aims to bridge this gap by providing a platform that helps agents effectively navigate documentation, API specifications, and best practices. The platform offers detailed tracing of I/O, skill invocation, and tool usage, which is instrumental in debugging and improving agent performance. Klingen noted that the Langfuse documentation itself, comprising 478 markdown files, underscores the complexity involved in creating comprehensive resources for AI development.

Lessons Learned in Skilling Up Agents

Klingen outlined six key lessons derived from their experience:

• 1. Qualitative Trace Analysis: Simply looking at traces can provide up to 80% of the understanding needed to debug an agent. Langfuse's interface allows for easy viewing of skill invocation positions, tool usage, and intermediate results, providing valuable context.
• 2. Capturing Production Signals: Observing real-world usage data, including queries and responses, can reveal critical patterns and identify areas where agents might be struggling or where new skills are needed.
• 3. Guiding Agent Navigation: Agents need to understand how to navigate available information effectively. This involves providing them with structured documentation and clear pathways to access relevant data.
• 4. Basic Evaluation is Better Than None: Even a simple evaluation setup can significantly boost confidence in iterating on agent skills. The team found that having a basic setup allowed them to release new versions with improvements and regressions identified more readily.
• 5. Referencing Dynamic Content: Instead of copying documentation, referencing it dynamically ensures that agents always have access to the latest information. This is crucial as documentation and APIs evolve.
• 6. Auto-research is Bounded by Target Function: The effectiveness of automated research for agent improvement is constrained by the defined target function. The team found that by focusing on specific tasks like improving prompt migration skills, their auto-research efforts yielded better results.

The Importance of Feedback Loops

Klingen emphasized the role of feedback loops in agent development. By tracing agent execution and analyzing the outcomes, developers can identify discrepancies and implement necessary improvements. This iterative process, supported by tools like Langfuse, is essential for building reliable and capable AI agents.

The presentation concluded with a look at the future, highlighting ongoing challenges such as skill upgrade lifecycles, skill distribution, and the precise definition of agent target functions. Klingen expressed excitement about the future of AI engineering and the potential for agents to become more sophisticated and integrated into various applications.