Understand the whole assistant first

An assistant receives a question, gathers information, proposes a next step and checks whether it can finish. In a workflow, those steps are chosen in advance. In an agent, a language model chooses some of the steps using the goal and previous results. A tool is an operation the surrounding program permits, such as reading an approved document.

Your first task is small: answer a policy question with an exact quotation and source label. Then decide whether that quotation actually answers the question. This gives you a complete loop to understand before adding memory, external services or more freedom.

1. Complete the existing first exercise

Open the source-reading guide and runnable starter. It provides one downloadable Python file, fictional policy notes, every run command, complete expected output and troubleshooting. You need Python 3.10 or later, a terminal and about 30 minutes. No account or API key is needed for the scripted replay.

The question asks when an unused item can be returned. The fictional source says within 14 calendar days of delivery. Your finished result is the source quotation, its label and a visible review step. The supplied program includes checks for invented quotations, unknown tools, unapproved sources and its action limit.

2. Separate scripted checks from model behavior

The replay makes fixed choices so every reader can compare the same output. It does not call or evaluate a model. The optional interactive exercise lets you copy a real model's proposed action into the tool program and carry the result back. That is a manual model-and-tool loop, not an automated deployment.

Try a question the sources cannot answer, such as whether opened items qualify. The useful response is to withhold an unsupported conclusion. A source label alone does not establish relevance or truth; the person reviewing the answer still has work to do.

3. Follow a task, rather than every edition number

The published lessons below build six parts of a useful assistant: a research brief, a read-only tool, memory, an approved file save, evaluation and recoverable document work. Choose the part relevant to your project. Each is a complete local exercise with its own limitations.

The next application should connect parts into a task a reader recognizes, such as preparing a brief from approved documents or drafting a client follow-up for review. State when ordinary search, a checklist or a fixed script would already do the job. Adding model choices is a design decision to test.

Where trading belongs

The journal keeps its original name, editions and research history. Trading remains a demanding specialist case study for questions about evidence, timing, execution and cost. It is not a prerequisite for learning to build assistants. Synthetic and paper experiments do not establish funded trading performance.

You are finished with the first step when you can run the starter, explain why an invented quote is rejected, identify an unanswered question and describe who reviews the result. If you need a complete business workflow with less programming, choose AI Operator Weekly.

For the underlying distinction, see Anthropic's Building effective agents. The journal's public examples supply their own inputs and instructions. Private repository access is never a learner prerequisite.