The pedagogy

Built on learning science, not framework hype.

You're not lazy and you're not bad at this. The problem is that most agentic-AI material asks you to read a framework's docs, run its demo, and hope the judgment transfers. It doesn't. AgentGraph replaces all of that with calibrated practice on the decisions that actually make an agent work. Here's exactly how.

The system

Four moving parts, working in tight feedback.

Every decision — what you see next, when a topic resurfaces, whether an answer counted — comes from these four components interacting.

STEP 01

A diagnostic calibrates where you are.

Your first session isn't a lecture — it's a short, adaptive diagnostic that probes which design calls you can already make. Within ~30 minutes, the system has a live map of which decisions you've mastered, which are shaky, and which are blanks. That map becomes the starting state of your graph.

STEP 02

A decision graph models every system and the calls that make it work.

Ten canonical agentic systems — a support agent, a RAG assistant, a coding agent, a deep-research agent, a multi-agent system, and more — each broken into the decisions that make or break it: autonomy level, context strategy, memory, tool design, model routing, control flow, evaluation. Every decision is a node; every prerequisite is an edge. The graph is what puts you on the exact frontier of what you can design next — never trivial, never impossible.

STEP 03

Each session is a sequence of calibrated drills.

No framework tours. No demos to watch. You open AgentGraph and you get 4–8 short drills — sometimes choosing an autonomy level, sometimes designing a tool interface, sometimes defending a trade-off in writing. Every drill targets a specific decision on your frontier. The grader reads your answer with the strictness of a Staff engineer, then gives you precise feedback — and the ideal answer — within seconds.

STEP 04

Spaced review keeps everything from fading.

Forgetting is not a character flaw — it's the default state of the human brain. The graph tracks when you last defended a decision and schedules it back into your review queue right before it would decay. Six weeks in, the call you made about context windows is still warm.

STEP 05

Explainers appear exactly when you need them.

When you get a drill wrong, the system doesn't tell you to go read a framework's docs. It surfaces a short explainer that opens with the concrete failure you just hit — the agent that cancelled the wrong order, the RAG answer that hallucinated — then puts you right back in the driver's seat with a fresh drill on the same idea.

STEP 06

Your dashboard tells you the truth.

Not "streak = 14 days". Not "you've watched 74% of the course". Your dashboard shows, for every decision on the graph: design-ready, shaky, or not yet attempted. That's the only metric that matters when you're about to design an agent for real.

The graph

This is roughly what your graph looks like.

Lit-up nodes are decisions you can defend. Glowing edges are prerequisites that unlocked something new. Dim nodes are what's coming. Every session re-shapes it.

See yours →
Learning principles

Five ideas we don't compromise on.

Principle 01

Designing beats reading.

Reading the ReAct paper is not designing an agent. Every minute you spend with AgentGraph is a minute you're actively making and defending a design decision.

Principle 02

Specific beats general.

Generic "agent best practices" don't carry you through a real design. We break every system down to concrete, testable decisions — and drill each one until your judgment is automatic.

Principle 03

Feedback beats answer keys.

Bad feedback: "correct / incorrect". Good feedback: "your agent can cancel any order because authorization lives in the prompt, not in a deterministic gate — so a single jailbreak bypasses it." That's what our grader produces.

Principle 04

Calibration beats content.

A pile of 200 agent tutorials doesn't help if you can't pick which decision to practice today. The graph picks for you — so you spend your 30 minutes building judgment, not choosing.

Principle 05

Retention beats coverage.

Touching every axis once is worse than internalizing the core ones three times at the right intervals. We pick retention every time.

Principle 06

Honesty beats hype.

No "you're 87% ready!" dopamine-bait numbers. If your judgment on autonomy and control is shaky, the dashboard says so. That's the whole point — design for real knowing what you actually know.

You've read enough. Open the graph.

30 minutes. No framework tours. You'll feel the difference by the end of session one.