Building agents

The body of an agent: what it says, how it runs, how data shaped like anything other than a string gets in and out.

        flowchart TB
    subgraph instruction[What the LLM sees]
        P[Prompts<br/>P namespace]
        C[Context<br/>C namespace]
    end
    subgraph runtime[How it runs]
        E[Execution<br/>.ask / .stream / .map]
        K[Callbacks<br/>before_*/after_*]
        R[Presets<br/>reusable bundles]
    end
    subgraph data[What flows through]
        S[State transforms<br/>S namespace]
        D[Structured data<br/>.returns Schema]
    end

    P --> E
    C --> E
    E --> K
    D --> S

    classDef inst fill:#fce4ec,stroke:#ad1457,color:#880e4f
    classDef run fill:#e8f5e9,stroke:#2e7d32,color:#1b5e20
    classDef dat fill:#fff3e0,stroke:#e65100,color:#bf360c
    class P,C inst
    class E,K,R run
    class S,D dat
    

Chapters

Chapter	Read it when
Prompts	You want role + task + constraints + examples in one place (P.role() + P.task() + …).
Execution	You need .ask() vs .ask_async(), streaming, sessions, or batched .map().
Callbacks	You need to audit, retry, or mutate inputs/outputs at before_*/after_* boundaries.
Presets	The same model/instruction/tool bundle is showing up in five different agents.
State transforms	You need to reshape state keys between steps (S.pick, S.rename, S.merge).
Structured data	You want the LLM to emit JSON that matches a Pydantic schema, not prose.
Context engineering	The agent is seeing too much (or too little) conversation history.

New here? Finish Foundations first — this tier assumes you know what a builder and a state key are.