Hello World, LangChain

A first LangChain chain, from prompt template to LCEL pipe, with a local model swap, response inspection, tests, type-checking, and tracing added around it.

Agent Loops in Python

What actually happens between an LLM call and a tool call: the read-act-observe loop that turns a plain language model into an agent.

Tool Calling with LangChain

How create_agent and the @tool decorator collapse a hand-rolled tool-calling loop into a few lines, and what that trade-off actually costs you.

Graphs, Not Chains

Why LangGraph models an agent as an explicit graph of nodes and edges instead of a hidden loop, and the core primitives that make that possible.

Durable State and Checkpoints

How LangGraph snapshots a graph’s state after every step, so a run can survive a crash, get inspected mid-flight, or resume exactly where it left off.

Human-in-the-Loop Interrupts

How LangGraph’s interrupt() pauses a running graph mid-step to ask a human before a risky action, and how Command(resume=…) picks it back up.

Structured Output as a Guardrail

How binding a Pydantic schema to a model or agent turns a probably-shaped response into a validated object, and what still isn’t guaranteed once it is.

A Minimal Eval Harness

A small, hand-rolled harness for checking whether an agent’s output is actually correct, not just well-formed, with a fixed dataset, a scorer per case, and a pass rate.

LLM-as-Judge

When correctness is subjective, grade agent output with a second, structured LLM call instead of eyeballing every run.

Tracing Agents with OpenTelemetry

When one specific agent run goes wrong, aggregate pass rates won’t tell you where. A span tree of every model call and tool call will, whether it’s built automatically, by a decorator, or by hand.

Timeouts, Retries, and Loop Limits

A timeout bounds a call, a retry survives a transient failure, and a hard iteration limit is what stops an agent that never planned to.