Agentic Workflow Patterns 3 · Delegation and critique, and what each costs

[akash-coded]

Exercise 3: Delegation and Critique

Est. time: 40 min | Difficulty: intermediate | Patterns: v5 orchestrator-workers, v6 evaluator-optimizer

Practice: MCQ, choose-an-option, fill-the-blank code, fix-code, complete-a-flowchart, table-fill, spot-the-error, trace-the-delegation, name-the-failure-mode, one-line-diff, refactor-to-cheaper, estimate-the-bill, two-truths-and-a-lie.

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

---

Scenario

The complaint pile grows and the apology letters pile beside it.

• Sub-scenario 1: Nadia's mess. "Cancelled my BLR-DEL, I'm Gold, I want a refund OR the next flight, and I paid for a seat, what happens to it?" One message, four tangled asks, no clean category.
• Sub-scenario 2: Sofia's bar. Every customer-facing apology must pass a written policy and tone check before it is sent. Sofia in compliance will not accept "close enough."

Two patterns, two problems. Do not mix them up.

---

Part A: Match each sub-scenario (MCQ)

Sub-scenario	a	b	c
Nadia's mess	Routing	Orchestrator-workers	Chaining
Sofia's bar	Parallelization	Evaluator-optimizer	Routing

---

Part B: Choose the design for Nadia

• Design 1: a router with a fixed classifier that picks one branch.
• Design 2: an orchestrator that consults specialists as tools and decides which ones at runtime.

Pick one. One line: what does routing fail to handle in Nadia's message? ________

---

Part C: Fill the blank (agents as tools)

Complete the orchestrator so it can delegate.

flight_specialist  = Agent(model=haiku, name="flight_specialist",  system_prompt="...", tools=[get_pnr])
fare_specialist    = Agent(model=haiku, name="fare_specialist",    system_prompt="...", tools=[get_pnr, get_fare_rules])
refund_specialist  = Agent(model=haiku, name="refund_specialist",  system_prompt="...", tools=[get_pnr, check_refund_eligibility])

orchestrator = Agent(
    model=haiku,
    name="orchestrator",
    system_prompt="Read the message, consult only the specialists needed, then synthesize one reply.",
    tools=[________, ________, ________],     # the three specialists
)

---

Part D: Fix the broken delegation

Meera's version throws no error, but the orchestrator cannot address one specialist by name.

flight = Agent(model=haiku, name="flight_specialist", system_prompt="...", tools=[get_pnr])
fare   = Agent(model=haiku, name="fare_specialist",   system_prompt="...", tools=[get_pnr, get_fare_rules])
refund = Agent(model=haiku, system_prompt="Decide refund eligibility.")    # <-- here

orchestrator = Agent(model=haiku, name="orchestrator",
                     system_prompt="Delegate to specialists.",
                     tools=[flight, fare, refund])

• What is wrong: ________
• The one-line fix: ________

---

Part E: Trace the delegation

Nadia's message hits the orchestrator. List a plausible order of specialist calls the model would make, and the one part of her message each call resolves.

Step	Specialist called	Part of Nadia's message it answers
1	________	________
2	________	________
3	________	________
4	________	________

Then answer: on a different message ("just tell me my gate"), how many specialists would the orchestrator call, and why? ________

---

Part F: Complete the evaluator flowchart

Fill the two edge labels.

flowchart TD
    In([Draft apology]) --> D[draft node]
    D --> C{critic node}
    C -->|... fill 1| D
    C -->|... fill 2| P[publish node]

Loading

• Label 1 = ________
• Label 2 = ________

---

Part G: Fill the edge table

For the cyclic evaluator graph, complete the condition column.

From	To	Condition
draft	critic	________
critic	draft	________
critic	publish	________

---

Part H: Spot the error and name the failure mode

Sofia's critic is strict and rarely satisfied on the first pass. Meera ships this graph.

b.add_edge("draft", "critic")
b.add_edge("critic", "draft",   condition=needs_revision)
b.add_edge("critic", "publish", condition=is_approved)
b.set_entry_point("draft")
graph = b.build()

• What is missing: ________
• Name the failure mode in three words or fewer: ________
• The one-line diff that stops it: ________
• The second line that keeps each redraft clean: ________

---

Part I: Red-team the critic

Ravi writes a critic prompt that says only "make it good."

• Why does this loop burn tokens without improving much: ________
• What must the critic prompt contain to be worth running: ________

---

Part J: Refactor to cheaper

Priya finds an orchestrator wired for a queue that splits cleanly into exactly four categories, one branch each, no overlap.

orchestrator = Agent(model=haiku, name="orchestrator",
                     system_prompt="Delegate to specialists.",
                     tools=[status_agent, change_agent, refund_agent, baggage_agent])

• What is over-engineered here: ________
• The smaller pattern that fits: ________
• One line: what you gain by switching: ________

---

Part K: Estimate the bill

Sofia's apology loop runs a generator plus a critic each pass. Assume a per-pass cost of $0.003.

• Cost if the draft passes on pass 1: ________
• Cost if it takes 4 passes: ________
• Cost with an uncapped loop and a critic that never approves: ________
• The cap you would set, and why that number: ________

---

Part L: Two truths and a lie

One is false. Mark and correct it.

1. Orchestrator-workers and routing draw the same tree, but the orchestrator lets the model choose how many branches run.
2. An evaluator-optimizer loop is the largest single lever on customer-facing quality.
3. A voting loop and an evaluator loop are the same pattern with different names.

---

Skeptic's corner

A teammate wants an evaluator-optimizer loop on every response the system produces, including internal status lookups.

• Where does the loop earn its cost?
• Where is it pure waste? Two lines.

[NimeshNairIBS]

<style> </style>

Part A
	Nadia's mess	Orchestrator-workers
	Sofia's bar	Evaluator-optimizer
PartB
	Design 1: a router with a fixed classifier that picks one branch.
Part C	flight_specialist  , fare_specialist , refund_specialist
Part D	What is wrong	no tools defined
	The one-line fixAgent(model=haiku, name="refund_specialist",  system_prompt="...", tools=[get_pnr, check_refund_eligibility])
PartE
	Step	Specialist called	Part of Nadia's message it answers	"Cancelled my BLR-DEL, I'm Gold, I want a refund OR the next flight, and I paid for a seat, what happens to it?" One message, four tangled asks, no clean category.
	1	flight_specialist	Next Flight Information
	2	fare_specialist	Flight cancellations and seat
	3	refund_specialist	Refund
	4	loyalty_specialist	Loyalty tier
	Then answer: on a different message ("just tell me my gate"), how many specialists would the orchestrator call, and why? ________
			1, Flight Specialist
PartF	Label 1	Revision Needed-Feedback
	Label 2	Approved
Part G: Fill the edge table
	For the cyclic evaluator graph, complete the condition column.
	From	To	Condition
	draft	critic	Revision
	critic	draft	Feedback
	critic	publish	Approved
Part H: Spot the error and name the failure mode	???
Part I: Red-team the critic
	Why does this loop burn tokens without improving much:	No exact feedback was provided
	What must the critic prompt contain to be worth running	Critic promp should contain the criterias which should be adhered to get it approved
Part J: Refactor to cheaper
Part K: Estimate the bill
	Cost if the draft passes on pass 1:	0.003
	Cost if it takes 4 passes	0.012
	Cost with an uncapped loop and a critic that never approves	Infinite
	The cap you would set, and why that number: _	atlesst 5.
Part L: Two truths and a lie
	1 is False
Skeptic's corner
	In the feedback loops
	needs to be used only if the quality response is non negotiable. Eg contract document etc.

[ksvivekanand86-source]

Part A:
Nadia's mess: b
Sofia's bar: b

Part B:
Design 2
Routing picks one branch but Nadia's message needs multiple specialists.

Part C:
tools=[flight_specialist, fare_specialist, refund_specialist]

Part D:
What is wrong: Missing name parameter on refund agent
Fix: refund = Agent(model=haiku, name="refund_specialist", system_prompt="Decide refund eligibility.")

Part E:

1. flight_specialist - confirm cancellation
2. refund_specialist - check refund eligibility
3. flight_specialist - find next flight
4. fare_specialist - check seat payment

For "just tell me my gate": One specialist (flight_specialist). Simple query, single concern.

Part F:
Label 1: needs revision
Label 2: approved

Part G:
draft → critic: always
critic → draft: needs revision
critic → publish: approved

Part H:
Missing: max iteration limit
Failure mode: infinite loop
Fix: graph = b.build(recursion_limit=5)
Second line: b.add_node("draft", draft_agent, state_checkpoint=True)

Part I:
Burns tokens: no specific criteria means unclear feedback and random revisions
Must contain: specific measurable criteria and clear pass/fail conditions

Part J:
Over-engineered: using orchestrator instead of simple routing pattern
Better pattern: routing
Gain: lower cost

Part K:
Pass 1: $0.003
4 passes: $0.012
Uncapped: infinite
Cap: 5 iterations

Part L:
Statement 3 is false.

[Vysakhan208853]

A
Nadia : b
Sofia : b

B
Design 2, because of non deterministic nature of input

C
flight_specialist, fare_specialist, refund_specialist

D
name is missing

E
flight_specialist - next flight?
refund_specialist - refund?
fare_specialist -

F
revision
good to publish

G
draft critic draft-ready
critic draft need-revision
critic publish looks-good

[Vysakhan208853]

H
infinite loop problem

I
"make it good." is vague feedback and loop burns
give specific feedback on what and where to improve

J
Orchestrator
classifier (rule based)
cost

K
$0.003
4 * $0.003
infinite
3, a good draft agent would not need more than 3 for drafting an apology

L
3

M
Not necessary for every response, specifically when retrieved data is static in nature