Saying it out loud

Lesson 5 of 6 in Whiteboarding Agent Systems: Drawing the Thing You Built.

A board is half the performance. The other half is a script, and the script has a shape: one sentence per box, in the order you drew them, and never a sentence about the notation. Nobody wants to hear that round boxes mean model calls; they want to hear what the system does.

Here is the whole narration for the supervisor sketch, timed at about forty seconds.

A worked script — forty seconds, sketch 2

“A customer exception arrives as mail — untrusted, so nothing in it is ever an instruction. [route] Code classifies it into one of five types with a small model, and anything under 0.8 confidence goes to a human instead of down a specialist path.

[supervisor] For a shortage, a supervisor briefs two workers. The brief is references and one exact question each — contract ids and the SKU, not a summary of the mail — because everything that gets re-summarised at a boundary loses fidelity.

[workers] One checks stock and lead time at three warehouses. The other reads the master agreement to find what remedy the customer is entitled to, and it has to return a clause citation, so we can check it.

[merge] They come back into a merge that is real code, not a concatenation: entitlement constrains which remedies are legal, availability constrains what we may promise, and coverage is a field — right now this run only heard from two of three warehouses, and the proposal downstream knows that.

[gate] Code compares the proposal’s cost to $250. Under it, we execute. Over it, a human approves the exact parameters that will run.

The whole thing is capped at eight seconds and thirty thousand tokens across all workers, and it stops when a remedy executes or when it escalates. About three thousand runs a day.”

Notice what is not in there: no framework names, no notation explanation, no “and then the AI decides”. Six sentences, each naming a box and the decision it owns. Also notice the last sentence — the stopping condition and the volume — because it answers the two questions most likely to be asked first.

Then the questions come, and they are almost always the same three. Answer them with the marker, not with your voice. Adding to the board makes you the person who understands the system; talking over a static board makes you the person defending one.

“Where does it stop?”

Point at the two exits you already drew and say both aloud: the natural stop (the model returns text; a remedy executes) and the hard stop (turn cap, spend cap, wall-clock). Then add the third if they push: the escalation exit. If you find yourself saying “it stops when it’s done”, you have discovered a real gap — say so plainly and write STOP: ? on the board. A named unknown reads as engineering judgment; a confident non-answer reads as inexperience.

“What happens when the ERP times out?”

Draw the failure arrow live. Where does it go — retry inside the remaining budget, proceed with partial coverage, degrade to a human, or abandon? Then say the consequence: “two of three warehouses is enough to answer, so we proceed and label the coverage; if the entitlement worker fails, there is no legal remedy to propose, so that one escalates.”

The general form of the answer that always works: name which branches are load-bearing and which are best-effort. Reviewers are not testing whether you handle failure; they are testing whether you have distinguished the failures that matter.

“What can it reach?”

This is the security question in disguise, and the overlay is the answer. Add the dashed enclosure, name what holds it (identity, network, sandbox), then count the trifecta legs out loud. If all three are present — and in most business systems they are — say so and name the containment: “it can read customer data and it can send mail, so the mail tool only sends to the address on the order record, never to an address extracted from the message body.”

Volunteering the leg count is the single strongest move available to you in a security review. It signals you did the analysis, and it sets the agenda on your terms.

“Why not just use a workflow?” (the interview question)

A fair question and often the right answer. Say the trade rather than defending the choice: “the remedy decision needs judgment over an entitlement clause we cannot enumerate, so that hop is a model call. Everything either side of it is code — routing, the merge, the threshold — because those are enumerable. If the clause set were small enough to encode, I would write the matrix and delete the agent.”

Interviewers are listening for whether you can locate the boundary between the two, and whether you would give up the agent when the boundary moves.

One drawing remains, and it is the one people reach for last and need most. During an incident the question is not what can the system do but what did it do, in what order, and what did it change. That is a different picture: time across the page, hops as bars, and a mark wherever the world changed.

The incident drawing: lanes, not boxes

  run 7f3a91          0s    1     2     3     4     5     6      [3h51m]

  classify            ==
  resolve               =
  wh-1                   ==
  wh-2                   ============ TIMEOUT
  wh-3                   ====
  contract worker        =============================
  merge                                               =
  propose                                              =====
  gate                                                      | SUSPEND
  execute                                                   //  ==== ! !
  reply                                                             ==

  world changed at:                                             ^    ^
                                          SH-77120 (recallable to 18:00)
                                          CM-3391  (credit $412 posted)

Three properties make this the right drawing for an incident and the wrong one for a review.

Time is the axis, so “step 7 ran twice” is visible as two bars — which no architecture diagram can show. Parallelism is visible, so the room can see immediately that the contract worker, not the loud ERP timeout, owns the critical path. And the world-changed marks are separate from the bars, with their downstream ids and their reversal windows, because the first decision in an incident is what has to be undone and by when.

Draw this one in under a minute and you have replaced fifteen minutes of people reading log lines to each other.

Key terms: lethal trifecta, containment, stopping condition, trajectory

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.