Agentic AI for the AI PM
What agents actually are — the loop, tools, memory, and planning — and the reliability, security, and economics that turn demos into products.
Teams win with agents by matching autonomy to stakes, making the work verifiable, and earning trust one tier at a time.
A chatbot answers. An agent acts. Give a language model tools, a goal, and permission to loop — observe, decide, act, check the result, repeat — and you get software that can do things: research a market, fix a bug, book the logistics, run the workflow. That's the promise. The practice is a young discipline. It has real engineering behind it, but also a thick layer of hype on top — invented protocol acronyms, "8-layer stacks," and demos that collapse on contact with production.
This module is the honest map. It teaches what an agent actually is, and the machinery that makes one work: tools, context, memory, planning. It covers what changes when agents multiply — multi-agent systems and the real protocol landscape. And it covers the three disciplines that separate shipped agents from viral demos: reliability, security, and unit economics. Every lesson ships a diagram. The module opens with a knowledge graph so you can see how the pieces connect before you study them one at a time.
The knowledge graph
Agentic AI isn't a layer cake — it's a loop with disciplines attached. Everything in this module hangs off this picture:
What it knows · lesson 3
How it acts · lesson 2
How it thinks · lesson 4
Why it's trusted · lessons 6–7
Why it's a business · lesson 8
Read it in three passes. The spine: knowledge flows into the model. The model drives the loop. The loop acts through tools, and tool results flow back as new knowledge. That cycle is the agent. The amplifiers: planning and reflection make each cycle smarter. Multi-agent patterns run many cycles at once. The disciplines: every cycle leaves a trace. Error analysis turns traces into evaluators. Guardrails police the loop inline, while evals gate it offline. Security bounds every tool. A human approves the irreversible. Economics decides whether any of it is a business.
The lessons
- What is an agent? — the loop, the autonomy spectrum from workflow to agent, and when not to build one.
- Tools & function calling — how a model acts on the world: schemas, MCP, sandboxes, and the craft of tool design.
- Context & memory — the context window as working memory, retrieval, persistent memory, and context engineering.
- Planning & reasoning — ReAct, plan-and-execute, reflection, and what actually makes agents smarter (and slower).
- Multi-agent systems & protocols — orchestrators, subagents, and an honest map of the protocol landscape.
- Reliability & evals — compounding error, trajectory evals, observability, and why "95% per step" isn't good enough.
- Safety, security & governance — prompt injection, least privilege, human-in-the-loop, and audit trails.
- Agentic AI as a product — the capstone: unit economics, pricing, agent UX, and deciding where agents actually pay.
Each lesson pairs the concept with a 🎯 For the AI PM briefing. It names the decision the concept should change, and the question it should make you ask. Each lesson also has a diagram to make it concrete. Builders who want to go deeper than concepts can construct this machinery hands-on in the Harness engineering track.
Connects to other tracks
- Tools & function calling reliability — the contract layer under the agent's tool calls.
- Knowledge graphs & LLMs — GraphRAG as durable agent memory.
- Flowable — orchestration by a shared engine vs. by autonomous agents.
- Product sense for AI products — the taste and trust judgment agent UX demands.
📌 Close out the module: Recap & real-world examples.