Agentic Workflow Patterns 2 · The three dials and the bill they produce

[akash-coded]

Exercise 2: The Three Dials and the Bill

Est. time: 25 min | Difficulty: building | Patterns: v3 routing, v4 parallelization, cost modeling

Practice: MCQ, dials table, cost math, rank-them, debug-the-harness, spot-the-errors, red-team-the-guard, estimate-the-bill, complete-a-flowchart, two-truths-and-a-lie.

Anchor booking: PNR JX48Q2, surname Rao, Gold tier, BLR-DEL cancelled by the airline.

Illustrative Haiku prices: input $1.00, output $5.00 per 1M tokens.

---

Scenario

Thursday, the surge. A flash fare sale collides with a weather advisory. Traffic triples. Karan, who owns the cloud bill, is now standing at your desk watching the Bedrock spend tick up in real time.

Two engineers have opinions:

• Ana wants to parallelize the change-request flow "to cut the bill."
• Ravi wants to wrap every decision in a voting loop "for quality."

Both are about to be half-wrong. Your job is to settle it with numbers.

---

Part A: Spot the misconception (MCQ)

Ana says running three independent checks in parallel will save money. What actually happens?

• a) Cost drops because the calls overlap.
• b) Cost stays about the same; only latency drops, because every call still runs.
• c) Cost drops because parallel calls are billed at a discount.
• d) Cost rises because parallel tokens cost more.

---

Part B: Fill the dials table

Mark each dial versus a single-agent baseline as lower, higher, or same. Same task throughout.

Pattern	Cost	Latency	Quality on a hard task
Routing (v3)	____	____	____
Parallelization, sectioning (v4)	____	____	____
Parallelization, voting (v4)	____	____	____

---

Part C: Do the cost math

One model call costs:

$$ \text{cost}_{\text{USD}} = \frac{T_{in}}{10^{6}} \cdot p_{in} + \frac{T_{out}}{10^{6}} \cdot p_{out} $$

Two ways to answer the same refund query:

Design	Call	$T_{in}$	$T_{out}$
Routing	classifier	500	20
Routing	refund specialist	1200	300
Mega-agent	single call	3000	320

Fill in:

• Routing classifier cost = ________
• Routing specialist cost = ________
• Routing total = ________
• Mega-agent total = ________
• Cheaper design: ________
• One sentence, why the design with more calls still wins: ________

---

Part D: Rank them by cost

Four designs for one change request, with token profiles. Compute each, then order them cheapest to priciest.

Design	Calls and tokens ($T_{in}$ / $T_{out}$)
A. Routing	classifier 500/20, specialist 1200/300
B. Mega generalist	one call 3000/320
C. Parallel sectioning	3 workers each 800/150, aggregator 900/200
D. Voting x3	three calls each 1200/300

Ranking (cheapest first): ________ < ________ < ________ < ________

---

Part E: Debug the cost helper

This should total the cost of every call in a ledger. It has two bugs.

def block_cost(ledger):
    total = 0
    for tier, usage in ledger:
        p_out, p_in = PRICES[tier]                                  # line A
        total += usage["input"] * p_in + usage["output"] * p_out    # line B
    return total

• Bug 1 (line A): ________
• Bug 2 (line B): ________
• Rewrite the two corrected lines: ________

---

Part F: Red-team the parallel code

Ana ships this for the surge.

async def gather_change(msg):
    return await asyncio.gather(
        asyncio.to_thread(metered, fare_agent, msg),
        asyncio.to_thread(metered, reaccom_agent, msg),
        asyncio.to_thread(metered, loyalty_agent, msg),
    )

It works on your laptop and melts under the surge.

• Name the two guards it is missing: ________
• What failure shows up first under 3x traffic: ________

---

Part G: Estimate the bill

At the surge peak, TravelMind handles 10,000 change-like queries a day.

• Routing at your Part C total, daily cost = ________
• Mega-agent at your Part C total, daily cost = ________
• Yearly difference (365 days) = ________
• Karan asks for one lever to cut cost without hurting quality. Name it: ________

---

Part H: Complete the cost-decision flowchart

Fill the three blanks with a concrete lever.

flowchart TD
    Goal([Cut cost]) --> A[... fill A: cheapest model per stage]
    Goal --> B[... fill B]
    Goal --> C[... fill C]

Loading

• A hint given. Fill B and C with two more cost levers from the deck: ________ , ________

---

Part I: Two truths and a lie

One is false. Mark it, rewrite it correctly.

1. Parallel sectioning cuts wall-clock to the slowest subtask, not the sum.
2. Voting raises confidence and always lowers cost.
3. Routing sends the heavy work down one branch, so it usually beats a mega-agent on tokens.

---

Skeptic's corner

Ravi wants voting on every reply "to be safe."

• When is voting genuinely worth 3x the calls?
• When is it just a bigger bill? Two lines.