I've been covering AI agents every week for over a year now. Time flies!
Hundreds of hours of research, and published 40+ posts. Dozens of conversations with builders, investors, and teams deploying agents in production. Hundreds of comments from readers who are building in this space.
And the number one thing people keep asking me is: "Where do I even start?"
Not "what's the latest news." Not "which framework is best." Just... where do I start understanding what's actually happening with AI agents?
So i built the thing i wish existed when i first started going deep on this space.
a free, continuously updated field guide to everything you need to understand AI agents. Built for smart people who don't have time for noise.
What's inside
The handbook is 8 chapters. Here's the quick version:
Chapter 1: What Agents Actually Are. Everyone's using the word "agent" and meaning different things. I cut through it.
The core loop (goal → plan → act → observe → evaluate), the 5-layer agent stack, and why autonomy is a spectrum. Most things called "agents" are actually workflow automation with a model attached. That's not a bad thing, but you should know the difference.
Chapter 2: How Agents Work Under the Hood. Technical intuition without requiring a CS degree. Reasoning and planning (still the hardest unsolved problem), memory (it's not one thing. state vs. memory is a distinction that matters), tool use, multi-agent systems, and guardrails. What I’ve learned is that the model matters, but the system around it matters even more.
Chapter 3: Inside the Loop. What actually happens when an agent runs. Step by step.
Chapter 4: When Agents Break. This is the chapter I'm most proud of and observe very keenly. Prompt injection, data exfiltration, the confident wrong answer problem (far more dangerous than crashes), why benchmarks lie, and my 95-to-70 rule. Maybe less.
Chapter 5: How to Think About Agents. Frameworks for cutting through the noise. The reliability gap, the economic question (agents save 30-50% of cost today, not 90%), and the closed loop advantage - why the companies that win will be the ones generating their own training data from deployed agents.
Chapter 6: The Landscape. Who's building what? I’ve curated a list of 60+ companies, frameworks and agents that anyone should be tracking closely .
Chapter 7: Getting Started. Practical on-ramps depending on whether you're a builder, investor, or operator.
Chapter 8: The Reading List. Curated further reading, tiered by depth.
Why I made this free
Two reasons.
First, this space moves too fast. The handbook is a living document. I update it continuously as new developments happen.
Second, I think the biggest barrier to progress in AI agents is not technical. It's conceptual. People are making bad decisions (investing in the wrong companies, building the wrong products, deploying the wrong architecture) because the mental models they're working with are broken. If I can fix the mental models, everything downstream gets better.
Three things that might surprise you
1. No one has solved planning. It's a bundle of failure modes, like decomposition failure, ordering failure, tool selection failure, verification failure, objective drift. The hardest unsolved problem in agents, and it's where most of the research effort is concentrated right now.
2. The gap between demos and production is enormous. Almost anything can be demoed. Making it work 1000 times in a row? That's where most companies are stuck. The teams that win build for the 70% reality, not the 95% demo.
3. The real moat is the closed loop. Deploy an agent. It handles real tasks. Successes and failures become training data. Over time, your agent gets better at exactly the things your users need. Very hard to replicate. Very hard to compete with once it's running.
Go read it
It's free. It’s easy to read, and you can get through it in a Starbucks coffee session. It’s really targeted at folks learning about AI agents in a more serious way for the first time, but also a great quick refresher if you’re already into agents.
If you find it useful, share it with someone who's trying to make sense of this space. That'll be a compliment for me.
— Teng
P.S. The handbook is a companion to this newsletter. I'll keep covering the weekly news and deep dives here. The handbook is the foundation, the "here's how to think about all of it" layer that ties everything together.
P.S. Know a builder or investor who’s too busy to track the agent space but too smart to miss the trends? Forward this to them. You’re helping us build the smartest Agentic community on the web.