The automation ladder · Choosing a rung

Pick the rung for three tasks

Place three agency tasks on the ladder, price one step up against the task's own sum, and decide whether one agent should do all three.

casemediumExercise · 4 min · 17 of 75

On this page

Three tasks, three different rungs, and a founder who would rather build one agent for all of them. A rung too low costs incidents; a rung too high costs something every month.

The situation

The agency has approved three builds for this quarter. Assume staff at $30 an hour, the founder at $100, a month of 4.33 weeks, and the build estimates from the ladder lessons: about 15 minutes on rung 1, 12 hours on rung 2, 20 hours with tests on rung 3.

New client setup. 3 signed contracts a month. Each creates a project in the project tool, a shared folder and a client channel: 20 minutes by hand, no judgment.
Categorise bank transactions. 150 a week from the bank feed, 1 minute each, $20 per mis-post. One model call per transaction; anything it is unsure of goes to the operations manager's review queue by Friday. The accounting API fails a few times a week, and bank data may pass through no vendor except the model provider.
Screen CVs. 150 applicants per role, about one role a month, an answer promised within three days. The founder approves borderline applicants in two sittings a day, and a bias check on 40 saved CVs must pass before any prompt change goes live.

The founder's proposal: "One agent with the tools for all three."

The question

Place each task on a rung, price the step up to the next rung for one of them, and decide on the founder's proposal.

A strong answer · Rubric · Where answers go wrong

Write your answer first, then open

A strong answer

Client setup: rung 1. Three actions, no state, no branch. By hand it is 3 × 20 minutes = 1 hour, $30 a month; a 15-minute chain, $25, pays back in under a month. Rung 2 at 12 hours, $1,200, would need 40 months: never.

Bank transactions: rung 2, self-hosted. It needs a loop over the day's feed, a model call per item, a branch on the model's confidence into the review queue, retries and an error branch on the accounting API, and a wait until Friday. Rung 1 fails twice over: the model's output decides a branch, and the chain would run in a vendor's cloud, where the bank data may not go. Rung 3's tests are not worth 8 more hours for rules that change once a year.

CV screen: rung 3. Its state spans three days, the founder's approval is an interrupt, and the bias check is a test that must run before a prompt change, which a canvas cannot enforce. The step up to rung 4 would let the model choose which parts of each CV to read, so two identical CVs could take different paths. The bias check could then accept only a set of outcomes per case, and could never show that every applicant was read the same way. That rules it out before any token is priced.

One agent for all three: no. It puts a $20 mis-post, a client's folder and a hiring decision behind one prompt, one pair of caps and one eval set, and none of the three has steps that cannot be listed.

What would change this: transaction rules changing monthly would move that task to rung 3.

Rubric

places each task on a rung and names the capability that rules out the rung below: state, a branch on model output, an interrupt, a test before change
prices at least one alternative rung with build hours against the task's monthly sum, the arithmetic shown
puts the CV screen on rung 3 for the checkpoint, the approval and the pre-change test, not for "complexity"
says what rung 4 would cost the CV screen: a varying path per applicant that the bias check cannot assert
rejects the single agent with a reason tied to error costs or caps
names the evidence that would move a task

Where answers go wrong

"All three on rung 2, so there is one tool to maintain." The client-setup chain then costs 12 hours to save $30 a month, 40 months to pay back. One tool is a convenience; the sums decide.

"The CV screen on rung 4, because reading a CV needs judgment." Judgment inside a step is one model call. The rung question is who picks the steps, and an agent picking them per applicant makes the bias check unprovable. Rung 4 and the harness prices this kind of step up.

"Bank transactions on rung 1, since it is only a feed." The model's confidence picks the branch, one of the signs in rung 1 that a chain has outgrown its rung, and the bank data cannot pass through another vendor's cloud.