Find anything
Search the whole guide
Search concepts, patterns, frameworks, benchmarks, failure modes, playbooks, glossary terms, and curated paths.
Orientation
The field gets easier when you separate loop, tools, state, and evals.
Agentic AI sounds slippery because people use one word for many things: a function-calling chatbot, a workflow graph, a browser operator, a coding worker, a multi-agent team, or an app platform. This guide keeps those layers separate.
Concept map
The primitives that make agents legible
Read these cards as the stable vocabulary. Frameworks change faster than these concepts.
Stack view
What sits around the model
Patterns
Architectures are control surfaces
Use the smallest pattern that exposes enough control, state, and verification for the task.
Interactive tool
Architecture builder
Pick a workload and risk surface. The recommendation changes with tool and memory scope.
Ecosystem
Frameworks and protocols are different layers
MCP and A2A are interoperability protocols. Frameworks are implementation choices.
Interactive tool
Framework chooser
Filter by the kind of control surface you need. Treat this as orientation, then run a one-day spike.
Reliability
Evaluate the trajectory, not only the final answer
Agent systems fail in the path: wrong tool, hidden state, missing approval, stale context, or brittle policy handling.
Interactive tool
Failure-mode simulator
Select a failure mode and see symptoms plus mitigations.
Interactive tool
Eval checklist generator
Choose the areas you want to harden. The generated checklist is intentionally concrete.
Practice
Grow from beginner to operator by building small inspectable systems
The fastest learning path is not a framework tour. It is one useful agent, one trace, one eval suite, then one controlled workflow.
Vocabulary
Terms worth making precise
Evidence
Sources and refresh metadata
Future agents should refresh fast-moving framework and protocol claims before changing recommendations.