Build agents on self-hosted LLMs

​Agents are becoming the centre of modern software. Instead of clicking through screens, users state what they want and an agent works out how to do it across your systems. This session shows how to build that kind of agent and run it entirely on models you host yourself, so your data never leaves your environment.

​We will walk through application development with a LangChain Deep Agent, powered by a self-hosted LLM served with vLLM and governed by LiteLLM. Then we will show how Cognitivo used the same approach in production.

​What you will learn

• ​How an agentic software architecture is emerging, and why it puts the agent at the centre of your stack

• ​How LangChain Deep Agents handles planning, memory, sub-agents and human approval for long-running tasks

• ​How to host your own LLMs with vLLM (inference) and LiteLLM (gateway, keys, budgets and routing)

• ​Lessons from the field: an agent-first CRM that replaced a SaaS CRM, and LLM Bench, our tool for document extraction and comparing models

• ​When fine-tuning is worth it, and when better prompts or retrieval will do

​Agenda (AEDT)

• ​4:00 pm: Welcome

• ​4:03 pm: The emergence of the agentic software architecture

• ​4:13 pm: LangChain Deep Agents

• ​4:25 pm: Hosting your own LLMs with vLLM and LiteLLM

• ​4:38 pm: Field case study: CRM agent and LLM Bench

• ​4:53 pm: Why fine-tune?

• ​5:00 pm: Q&A

• ​5:30 pm: Close

​Who should attend

​Technology leaders, architects and engineers exploring AI agents or private LLM hosting. Some familiarity with APIs and LLMs helps, but no prior agent experience is needed.

​Hosts: Alan Hsiao, Aadhar Bhardwaj and Shaurya Sharma, Cognitivo