Build, monitor and debug multi-agent systems using Weights & Biases
Transitioning from simple LLM prompts to autonomous multi-agent systems introduces new hurdles, from unpredictable costs to "black box" behaviour. To move these systems into production, developers need more than just good prompts—they need a framework for evaluation and observability.
In this session, we’ll explore how to build, monitor, and improve agentic applications using Weights & Biases Weave. We’ll cover key strategies for maintaining high-quality AI, including:
Observability: Using traces to track token costs and performance across complex agent workflows.
Evaluation: Implementing "LLM-as-a-judge" to automate quality checks for helpfulness and brand safety.
Session Monitoring: Analysing multi-turn customer interactions to identify where agents succeed or fail.
To ground these concepts, we will walk through a contact centre modernisation use case for a fictional bank. You’ll see how specialised agents collaborate to solve real customer issues, like unblocking flagged transactions and increasing credit limits, across voice and chat. Join us to learn how to transform unpredictable agents into reliable, scalable business tools.
Connect with fellow AI engineers over refreshments and snacks during the networking after the presentation.
Please note: This is an on-site side event at Techarena 2026 that requires a ticket to enter Strawberry Arena. The side event is independently organised and run by the host of the side event. Please direct any questions (or praise!) to the event organisers.