# Agentic PDLC Exercises: 12 Problems with Worked Answers
Reading a formula and using it with a room waiting are different skills. These are the calculations the work actually asks for, in the order it asks.
9 min read · Intermediate · Lesson 4 of 4 in Practice · Updated 2026-09-24 · By Akash Das
Canonical: https://akash-coded.github.io/aws-bedrock-agentcore-strands/learn/agentic-pdlc-exercises/
> **In short.** Each of these twelve problems, three per phase, is a decision the agentic PDLC asks for, with the
> arithmetic that settles it, a worked answer and a link to the lesson that teaches it. The early ones
> ask whether work is AI work, what it is worth, which decisions halt the build, what the bar is and
> what chaining costs. The later ones ask whether a score proves a bar, how many cases and days proof
> takes, how long review waits, what the bill will be, when drift fires and what counts as a control.
**In this lesson** you'll practise:
- the four decisions of P0 and P1 that fix a project's shape and its bar;
- the proof arithmetic of P2: lower bounds, cases needed, days of evidence, review queues;
- the production arithmetic of P3: the bill, drift and what counts as a control.
## Sound familiar?
- You have read that a score needs a lower bound, and could not say whether 45 out of 50 proves 85%.
- The formula looked obvious in the lesson and slipped away in the meeting.
- Most AI courses test what you remember; the work asks you to calculate.
These twelve are the calculations the work asks for, each small enough to do on a phone.
## How should you use these exercises?
**Commit to an answer before you open the worked one.** Each takes two to five minutes with a
calculator, and each links to the lesson that teaches it. The numbers are hypothetical, chosen to make one
point each; the lower bounds use 1.96, with the Wilson bound under about 100 cases, as the rest of this
tutorial does.
## P0 · Frame
### Exercise 1 · AI work, or a rule?
Three candidate features: **(a)** route each insurance claim to one of four teams by the product code on
the form; **(b)** draft replies to 800 complaints a day, reviewed before sending; **(c)** approve wire
transfers above $10,000. **Which is AI work?**
Show the answer
**Only (b).** (a) has no judgement call: two competent people would route the same code the same way,
so it is a rule. (b) has judgement, volume and a recoverable mistake, because a draft is reviewed before
it goes. (c) may involve judgement, but a wrong transfer is not recoverable, so at most it is assisted:
the model can prepare the case, and a named person approves every transfer.
Show the answer
**$2,472 a day, rising to $2,664.** Gross: 300 × 12 × $0.80 = $2,880. Run cost: 300 × $0.40 = $120.
Review: 300 × 30% × 4 × $0.80 = $288. Net: $2,880 − $120 − $288 = $2,472. At 10% reviewed the review
line is $96, so the net is $2,664, which is why the review share belongs in the value line from day one.
Show the answer
**(b) is hard; (a) is soft.** Ask the four questions (reversible cheaply, a placeholder, an owner and a
date, downstream survives a change) and one "no" makes it hard. (a) passes all four. (b) fails the
first three: the decision cannot be cheaply reversed once refunds are being paid, "the model behaves
sensibly" is not a placeholder, and nobody owns it with a date. Settle it before the gate.
Show the answer
**90%, falling to 75%.** The bar is damage ÷ (damage + saving): 45 ÷ 50 = 90%. With the check, 15 ÷ 20 =
75%. A human hold on the risky step lowers the damage, and therefore the bar, which is how a feature
ships safely at a score it could never have reached unaided.
Show the answer
**73.5%, and replacing a step wins.** 0.95⁶ = 0.735. Raising one step to 99% gives 0.95⁵ × 0.99 = 0.766;
replacing it with code that is right every time gives 0.95⁵ = 0.774. Every step that can be exact should
be, before anyone tunes a prompt.
Show the answer
**Twenty-one.** Seven agents have 7 × 6 ÷ 2 = 21 possible pairings, and each is a place where context is
lost and a limit must hold. Start with one agent and tools; add an agent only for a named limit (a
context overflow, or parallel work a tool cannot express) and give its hand-off a named limit of its own.
Show the answer
**Only A.** A: 0.92 − 1.96 × √(0.92 × 0.08 ÷ 200) = 0.882, above 85%. B has under 100 cases, so use the
Wilson bound: 78.6%, below 85%. Its score is five points above the bar and proves nothing yet. A score
without its sample size has said nothing.
Show the answer
**About 514 cases, or 18 days.** Cases needed = 1.96² × 0.86 × 0.14 ÷ 0.03² ≈ 514. The slice sees
15% × 200 = 30 cases a day, and 514 ÷ 30 ≈ 17.1, so 18 days. A score closer to its bar needs many more
cases, because the gap is squared.
Show the answer
**4.8 days, falling to 2.2.** Two readers each: 24 slots ÷ 5 = 4.8 days. By band (two readers for R4,
one for R2 to R3, the harness alone for R1) 3 × 2 + 5 × 1 + 4 × 0 = 11 slots ÷ 5 = 2.2 days, with the money
changes read first and by more people than before.
Show the answer
**About $6,550; fix the routing, then the context, then the cache, then the breaker.** 1.5 × 1.4 × 1.2 ×
1.3 = 3.28, and $2,000 × 3.28 ≈ $6,550. Priority = (factor − 1) ÷ days: routing 0.8, context 0.5, cache
0.2, breaker 0.15. The breaker comes last and still matters: it is the fix that makes a runaway impossible.
Show the answer
**The baseline alert, in week 4.** The weekly change is 1.5 points, so the weekly alert never fires. The
level reaches 22.5% in week 3 (4.5 points) and 24% in week 4 (6 points), crossing 5. The breach should
re-open the release gate by itself, not open a ticket someone may read.
Show the answer
**Only (c).** (a) names a person, which changes nothing the system does. (b) is a request, and a model can
be talked past a request. (c) is enforced: the tool cannot send without the token, whatever the model is
told. Ask which enforced control would have made the incident impossible, and accept only code.