The subject is easier when you learn it as a dependency chain. Tools make little sense before structured output. Agent memory makes little sense before context. MCP makes little sense before tool boundaries and authorization.
If language models are new to you
Read these pages in order:
- How language models generate text
- Tokens and context windows
- Transformers and attention
- Sampling and structured output
- Embeddings and semantic search
- Retrieval-augmented generation
- What makes a system an agent
After that foundation, follow the backend engineer path.
If you build backend services
Start with the boundary between model output and application authority:
- Tools and function calls
- Safe tool execution
- The agent loop
- Context engineering and memory
- Durable execution
- MCP mental model
- Agent security
- Production readiness
The ideas should feel familiar. Idempotency, scoped credentials, queues, checkpoints, and tracing do not stop mattering because an LLM chose the next operation.
If you design agent platforms
Concentrate on decisions that affect many teams:
- Planning and control
- Agent orchestration
- Multi-agent systems
- MCP transports and lifecycle
- MCP security and operations
- Reliability
- Observability
- Evaluation
- Performance and cost
Then compare the design walkthroughs. They use the same concepts under different risk and latency constraints.
If you need a working example
The tutorials are deliberately small:
Each tutorial has one visible result. The explanation pages contain the deeper reasoning, so the tutorials can stay focused on the build.
If you are preparing for a design discussion
Read the architecture decision guide first. Pick one walkthrough and explain it aloud:
A useful explanation names the trust boundary, the source of truth, the stop condition, the recovery point, and the evidence used to judge success.