Building your own Agent harness

Most "agent" frameworks hide the parts that actually matter: what context gets sent to the LLM, how it picks the next action, and what happens when it gets something wrong. In this session, we'll look at why building your own harness gives you back that control, and how to do it without reinventing the wheel. You'll see how a thin, self-owned layer around the LLM makes your agents more predictable, cheaper to run, and far easier to debug.
What you'll learn:

• How to control exactly what context and tools get exposed to the LLM at each step, instead of relying on a framework's defaults
• Patterns for how the agent decides its next action (planning, routing, tool selection) and where to add guardrails
• Techniques to improve reliability: validation, retries, fallbacks, and handling bad or unexpected LLM outputs
• How to track and optimize token usage and cost across multi-step agent runs
• When it's worth building your own harness versus using an off-the-shelf agent framework

Will show how a real agent is structured in practice and how skills help solve common problems in execution and consistency, ending with a live demo on top of it we can have QnA

Presenter: Anukool Chaturvedi

Anukool Chaturvedi is working as a Sr. Machine Learning Engineer at Avalara, he loves to explore, learn, and build solutions incorporating AI, recently he has been building distributed Multi Agent solutions for many use cases, experimenting with various frameworks like LangGraph, Google ADK, CrewAI etc. He also write technical blogs. In his free time, he loves reading books on biographies, businesses and history.

He can be reached at -
Linkedin: https://in.linkedin.com/in/anukool-chaturvedi

Note : The speakers' views are their own and do not represent Avalara. This session is intended for educational purposes only and does not constitute specific advice or endorsement from Avalara or its organiser ! Additionally This meet-up event does not include any Avalara specific contents.