Building agents with Claude and LangChain

In 2023 I founded Tilegra: a conversational AI and omnichannel automation platform for business clients. The hypothesis was clear: agents weren't demos anymore — they worked in production if you built them right.

The stack

After a year of iterating, the heart of Tilegra is:

• Claude API as the main model — the right balance of quality and latency for our use case
• RAG with LlamaIndex + pgvector / Supabase so each agent has the exact context its client tells it to have
• Custom tool-use so the agent doesn't just answer but does things: hits APIs, sends emails, schedules
• Dynamic memory per conversation for real continuity
• Embeddable widget + webhooks with a TypeScript / Node.js / Fastify backend

From n8n to LangGraph

I built the first prototype with n8n. It worked to validate the idea but didn't scale: hard to version, hard to test, poor observability. I'm migrating everything to LangGraph with a code-first multi-agent architecture and full tracing in LangSmith.

An agent without observability is black magic. And magic doesn't scale.

OlivAI: the other side

In parallel, at Aardvark Partners I lead the development of OlivAI: a customer service agent for the Aardvark Labs suite. It's integrated with the internal Check-In platform, so when a guest writes in, the human agent already has all the context on hand — reservations, payments, prior communications.

What I learned

• That evaluation is 50% of the work — an agent without an eval set is a ticking time bomb
• That iteration speed on agents is the most underrated competitive asset of 2026
• That multi-agent ≠ better: sometimes a single well-built agent beats a poorly tuned orchestra
• That being truly AI-native (Claude + Claude Code + Cursor every day) changes how you decide the architecture