Skip to content
Expedify
Start free
15 min

Capstone: the desk, end to end

Build the whole thing in your own org — with a self-check made entirely of failures this course actually hit.

Build the whole desk. Not from a template — from the four parts in lesson 1, in your own org, against your own records.

There is no answer key for this one, and that is deliberate. Every worked example so far has been ours; this one has to be yours, because the failures you will meet are the ones your data has and mine does not.

The brief

An enquiry arrives from someone you have never heard of. By the time it is finished, that person should be: answered, recorded once, at a stage that is true, and either handled or handed to a named human with a deadline. Nothing you cannot defend should have been said, and nothing irreversible should have happened without a check.

Eight pieces. You have built each one in isolation.

The enquiry agent

From
L2, L3, L4
It is done when
It quotes a fee verbatim, and refuses one it does not hold.

Memory

From
L7
It is done when
It resolves "the second one" without carrying a figure forward.

The trace

From
L8
It is done when
The same person writing twice produces one contact, not two.

The stage

From
L9
It is done when
It moves on quoted evidence and refuses to move on enthusiasm.

The handoff

From
L10
It is done when
The task has an owner, a due date, and the record attached.

The chase

From
L12
It is done when
It does not chase anyone who has already paid.

The watch

From
L13
It is done when
It finds someone who paid and never started.

The digest

From
L14
It is done when
A buyer receives their own rows and nobody else's.

The self-check — and where to be suspicious

Every question below is a failure that happened while this course was being built. None is hypothetical.

  1. Does your knowledge base contain the facts your rules refer to? Ours had a policy instructing the agent to quote fees exactly, and no fees anywhere in it. Ask it your customers' most common question before you trust it.
  2. Have you watched every gate go BOTH ways? A condition observed only false is not tested. One of our gates referenced a field that did not exist and was false forever; the workflow looked perfect and did nothing.
  3. Did you check a zero is really a zero? A failed query and an empty result look identical from inside a workflow — same success flag, same empty data, no error.
  4. Does anything downstream reference a node that is a TOOL? Tools run inside their agent and never appear as top-level steps, so the reference resolves to nothing. Ours took weeks to find, because it sat behind an approval gate nothing automated could reach.
  5. Is any identifier coming from the model? Told to produce a record id in the right format, a model will produce a perfectly-formed id that matches no record. Identifiers come from queries.
  6. Did you read the schema, or assume it? Contacts have no name and no status. Two of our runs failed on that before writing a single row.
  7. Have you pointed every send at yourself first? For a fortnight. Read what actually arrives before a customer does.

The pattern in that list. Almost none of those are the model being wrong. They are the instruction being right and the fact being absent — a rule pointing at a price that does not exist, a reference pointing at a node that is not there, a gate pointing at a field that was never published. Check what your instructions point at.

What good looks like

You are finished when you can answer three questions about any conversation your desk handled last week, without opening the transcript:

  1. What did it tell them, and where did that come from?
  2. What changed in the CRM as a result?
  3. If it stopped, who picked it up, and when?

If all three are answerable from the records alone, you have a desk. If any needs the transcript, you still have a chatbot with a database attached.

Where to go next

An agent that can act on money, credentials or anything irreversible

Go to
Building Production-Ready AI Agents — the control layer, agent teams, the approval gate and the audit row.

To understand why the prompt behaves as it does

Go to
From Prompt to Harness — prompting, context engineering, loops.

To know whether any of it is actually working

Go to
Does It Actually Work? — traces, evaluators, golden sets.

One last thing, and it is the honest summary of the whole path. Nothing here made an agent smarter. Every lesson made one narrower — gave it fewer things it was allowed to say, fewer values it was allowed to author, fewer paths it could take without a human. That is what made it safe enough to leave running, and it is the same idea the flagship spends twenty-one lessons on.

Try this in Expedify →

Related lessons

12 min read

Base rates — what a piece of evidence is actually worth

A face-recognition system that is 99.9% accurate and almost entirely wrong, and a number that sent an innocent woman to prison. Both are the same arithmetic, and it is the arithmetic that decides what any piece of evidence is worth.

Read
10 min read

Confirmation and survivorship — what you never looked for

Two questions about evidence you did not go looking for. One is a rule you have to discover, and one is a pattern in five famous people — and in both, the thing that would have told you the truth is the thing nobody checks.

Read
14 min read

Loss aversion, sunk cost and regression — what it costs you

Four questions you answer about yourself rather than about a scenario, and your own answers are the finding. Then the pattern that makes praise look useless and criticism look like it works, whatever you actually do.

Read