Agentic AI — Multi-Agent & MCP Projects

Overview

A deliberately repetitive body of work: build the same class of agent system in several frameworks, then rebuild the interesting parts without a framework at all. The goal was never a portfolio piece — it was to find out which abstractions earn their weight once an agent has to run unattended.

What the projects cover

• Multi-agent teams with explicit planner, worker and reviewer roles, and real handoffs between them rather than one prompt pretending to be a team.
• Custom MCP servers exposing typed tools to agents, so capability is declared and discoverable instead of hard-coded into a prompt.
• RAG pipelines end to end: chunking strategies, embeddings, vector search, and reranking — plus the measurement to tell whether any of it helped.
• Structured outputs and guardrails, because an agent that returns free-form prose is an agent you cannot build on.
• Evaluation harnesses that score behaviour across runs, which is the only way to notice a prompt change made things quietly worse.
• Fine-tuning with QLoRA and deployment of the resulting LLM services.

What I took away

Three things, repeatedly:

1. Retrieval quality dominates. Most "the model is bad" symptoms were retrieval problems wearing a costume — wrong chunk size, no reranking, or an embedding that flattened the distinction the query depended on.
2. Frameworks are great until the failure path. Orchestration libraries make the happy path elegant. Retries, partial failure, timeouts, and resuming a half-finished run are where you end up reading the source.
3. Evals before features. Without a scoring harness, every prompt tweak is a vibe. With one, you can refactor aggressively and know within minutes whether you broke something.

Stack

Python throughout, with OpenAI Agents SDK, LangGraph, CrewAI and MCP for orchestration and tool exposure, and vector search for retrieval.