Modern teams are under pressure to ship LLM agents that actually work in production, but it's difficult to cut through the noise and determine what actually works. There's an ever-growing number of "agent frameworks" promising great results, yet whose abstractions are opaque and difficult to optimize. Blog posts and one-off repos explain pieces of the puzzle, but AI is moving faster than ever.

Many engineers struggle to:

- Choose the right agent pattern for their use case

- Incorporate reliable tool use into agentic workflows.

- Evaluate where and why agents fail.

- Deploy agents which optimize intelligence, cost, and latency.

- Understand when and how to improve agent performance with finetuning and RL.

We keep hearing that 2025 is the Year of the Agents. Everyone’s talking about MCP and A2A and GRPO but no one seems to agree on when you should use them. Agentic interactions are becoming table-stakes consumer features, and investors are eager to see that you’re keeping up with the times.

Popular agent products like Deep Research, Devin, and Manus are built by companies who don’t want to share their tricks. Open-source alternatives often underperform or are complex to understand and adapt. Textbooks don’t exist yet, and sifting through every new paper is basically a full-time job. The latest API models can make for powerful agents, but costs get out of control quickly. Few people outside of the big AI labs have hands-on expertise in optimizing LLM agents using reinforcement learning. Will and Kyle happen to be two of them.

---

What to expect:

Beyond core principles, this course emphasizes hands-on practice for building production-ready agents, including:

- How to integrate MCP tools for popular services like Notion, Linear, and Slack into your agent applications

- How to build your own MCP servers for custom APIs and data

- How to scaffold and prompt agents for complex tool workflows

- How to evaluate and interactively refine agents with human-in-the-loop prompting

- How to use rule-based and LLM-based evaluations as reward signals for RL or synthetic data filtering

- How to train cost-effective agents which outperform models like o3 at a fraction of the cost using GRPO

The course will have 2x weekly lectures for 3 weeks, and we will have additional sections for office hours (see schedule below). Lecture videos will be available to watch asynchronously, and we'll also have a Discord chat for offline discussions.

Lectures will incorporate live coding/prompting with tools like Cursor, Claude Code, and Jupyter notebooks. Familiarity with Python, high-level AI/ML concepts, and LLM APIs is assumed.

You will also receive:

- $100 in Prime Intellect compute credits

- $100 in OpenPipe finetuning credits

- 1 year of Weights & Biases Pro ($600 value)

---

Course schedule:

Lecture 1 (6/17)

Agent Patterns and Principles

- ReAct, MemGPT, Agentic RAG, Multi-Agent (A2A)

- Hands-on demos with HF smolagents + other frameworks

Lecture 2 (6/19)

Model Context Protocol: When and Why

- Client/Server architectures for tool calls

- Approaches to auth

- Hands-on agentic MCP flow demos with Claude Desktop + Claude Code etc.

Lecture 3 (6/24)

Evals for Agents

- Extending eval techniques to agentic workflows

- Rule-based vs LLM-as-judge

- Filtering rollouts for synthetic data collection

- Brief demo of SFT on filtered rollouts

Lecture 4 (6/26)

Reinforcement Learning for Busy Engineers

- Crash course in RL fundamentals without the math

- GRPO vs DPO vs PPO

- Demo of GRPO for training a reasoning model (via HF TRL)

Lecture 5 (6/24)

Formulating Business Problems as RL Tasks

- How to think about reward/rubric design for real-world tasks

- Environment = Tasks + Tools + Verifiers

- Walkthrough of problem formulation for email search (via ART)

Lecture 6 (6/24)

Training Agents with GRPO

- Deep dive into RL experimentation for agent workflows (via ART)

- Broader ecosystem: other RL trainers + integrations with existing agent/tool frameworks (smolagents, MCP)