Why a second part

Part 1 ends with a governed agent in seven files. Every idea that matters is in there: the loop, the tool, the guardrail, the gate, the log.

Part 2 is the same agent, rebuilt as something a team could actually run. Not because Part 1 was wrong — because four things change the moment a script has to survive being used.

The four changes

Seven scripts become one package. step7.py is a teaching artefact: one file you read top to bottom. Part 2 is run.py plus a fraud_triage/ package, because a system with five stages and eight tools does not fit in a file you can hold in your head.

One loop becomes a pipeline. Part 1 has a single agent that calls tools until it is finished. Part 2 has five named stages — Intake → Enrichment → Risk → Case → Human gate — and only two of them involve a model at all. Naming the stages is what makes the system reviewable: you can point at the one that went wrong.

Judgment leaves the model entirely. Part 1 argued that a hard rule belongs in code. Part 2 does it: the risk score is computed by explicit weights in rules.py, and no model touches the number. Run the system ten times and the scores do not move.

Defensive plumbing appears. Real tool calls arrive malformed, or aimed at the wrong customer. Part 2 has a function whose only job is to throw away the model’s arguments and substitute the right ones. Part 1 never had to pay that tax.

What you should already have

Part 2 assumes Part 1 — the loop, structured output, the difference between an instruction and a guardrail. It also assumes the same setup: uv, Ollama, qwen2.5:3b. Nothing new to install.

The result you’re building toward

ALERT-001  ->  ESCALATE  (risk 100/100, confidence high)
ALERT-002  ->  CLEAR     (risk   0/100, confidence high)
ALERT-003  ->  MONITOR   (risk  35/100, confidence low)

Those three numbers are the same every single run. The wording of the case summaries is different every time, and so is the set of tools the model chooses to call. Holding both of those facts at once is the whole point of Part 2.

How to follow along

The code is in the course repo under part2/, and its git history is the syllabus — each step below is one commit, and the code runs at every commit:

git clone https://github.com/nhial/building-ai-agents
cd building-ai-agents
git log --oneline -- part2/     # the eight steps
git switch --detach <sha>       # jump to any one of them
cd part2 && uv run --with ollama run.py

You can read the pages without doing that. But being able to run step 4 exactly as it was, then step 5, and diff them, is worth more than any explanation on this site.