Session 8. Loops and graphs — Wed 23 Sep

Conclusion

You ran one task three ways and filled the comparison table from counted calls: the chain spends one, the loop one plus a corrective retry, reflection three with the revision capped at one.

What you did

• Two of those columns are facts the code answers, and ch08-e1 re-read the code rather than taking your word for the numbers.
• You either ran the graph appendix and named what its trace showed, or you said why you did not — ch08-e2 takes both and refuses a blank.
• Then you wrote the workflow as a graph without a framework: five states, six edges in a dict, and step(state, event) as the only thing that moves between them.
• ch08-e3 drove your machine over every state and every event, legal and not, with no model anywhere in it.

The failure you handled

An invalid transition. An event that is not declared from the current state leaves the state exactly where it was, raises nothing, and comes back with rejected naming the event and the state it was refused from.

You injected the other two failures and read the difference. A retry storm is an edge with nothing counting on it, and visited, which only grows, is the counter that caps it. A partial tool failure is a decision about which event to emit, made in the open, with the failed source named and the flag travelling with the answer.

What to carry forward

done is terminal for one reason only — no row in the table starts there — so terminality is a fact about the data and not a guard somebody can delete.

Take it further

Today's graph has no model inside it, on purpose. Project 03 — Did the extra agents earn their calls? puts four of them in one: a coordinator that picks the company without calling a model, a researcher with read-only tools, a writer that may cite only what came back, and a critic that gets exactly one revision. You build it twice, once with LangGraph and once in plain Python, and you count what the second, third and fourth role cost against a single loop answering the same question. Optional, never counted.

Into session 9

Session 9 assumes you can draw the graph your agent walks, and it starts measuring whether that graph is the right one — structure was today; quality is the golden set and the pass rate.