Production Recap 4 · The lifecycle, the gates and the ambition ladder

[akash-coded]

Exercise 4: The lifecycle and building the agent

Language: Python, concept, diagrams Topics: P0 to P3, gates, ambition ladder, doors, tools as contracts, the loop and its guard Level: applied (debugging enters here)

Fourth foundation. Reading and spotting are assumed now; you also find and fix bugs. Each fix is one specific change, given as an option.

Q1. A team ships to P3 with no eval suite and no instrumentation. Which gate did they fail, and the named consequence?

• A) P2 to P3, they have nothing to test against
• B) P0 to P1, they built something that never earns out
• C) P1 to P2, they built with no acceptance bar
• D) P3 to operate, they shipped without a sign-off

Show answer

A) A supervised MVP with instrumentation and an eval suite is the P2-to-P3 gate. Skip it and P3 cannot begin.

Q2. The agent can (i) show rebooking options and (ii) confirm a rebooking that charges a fare difference. Autonomy differs because:

• A) both are irreversible, so both need human approval
• B) showing options is reversible and can run automatically; a charge is hard to undo and needs approval
• C) confirming can be automated once it is well tested; showing options is the part that actually risks misinformation
• D) both are safe to automate once the agent is well tested

Show answer

B) A two-way door can run on its own. A charge is a one-way door, so it needs a human. The axis is reversibility, not test coverage.

Q3. This loop has one defect that makes it unsafe before QA even starts. The fix is:

def run_agent(msg):
    messages = [{"role": "user", "content": [{"text": msg}]}]
    while True:
        resp = client.converse(modelId=MODEL, messages=messages, toolConfig=TOOL_CONFIG)
        messages.append(resp["output"]["message"])
        if resp["stopReason"] != "tool_use":
            return resp["output"]["message"]["content"][0]["text"]
        block = tool_use_block(resp)
        result = TOOLS[block["name"]](**block["input"])
        messages.append({"role": "user",
                         "content": [{"toolResult": {"toolUseId": block["toolUseId"],
                                                     "content": [{"json": result}]}}]})

• A) the toolResult is missing its toolUseId, so the model cannot bind it
• B) the assistant message is appended before the tool runs, corrupting order
• C) while True has no turn cap, so a mis-stepping model loops without end
• D) stopReason should be compared with end_turn, not tool_use

Show answer

C) The id is present and the ordering is fine. The missing piece is a bound like for turn in range(max_turns). No stop condition is a runaway.

Q4. This loop returns the right answer for a Gold passenger but is still broken. The bug is:

    for _ in range(6):
        resp = client.converse(modelId=MODEL, messages=messages, toolConfig=TOOL_CONFIG)
        messages.append(resp["output"]["message"])
        if resp["stopReason"] != "tool_use":
            return resp["output"]["message"]["content"][0]["text"]
        block = tool_use_block(resp)
        result = TOOLS[block["name"]](**block["input"])
        messages.append({"role": "user",
                         "content": [{"toolResult": {"toolUseId": "tool",
                                                     "content": [{"json": result}]}}]})

• A) the turn cap of 6 is too low for three tool calls
• B) the tool result should be appended with role assistant, not user
• C) TOOLS[block["name"]] should be handed the whole tool-use block object, not only its parsed input dictionary
• D) toolUseId is hardcoded to "tool" instead of block["toolUseId"], so results do not bind to their calls

Show answer

D) A fixed id happens to work with one call in flight, then binds results to the wrong call. Use block["toolUseId"].

Q5. Complete the agent loop. Bank: a) has enough b) needs data c) on error d) on approval

flowchart LR
  U["User message"] --> M["Model decides"]
  M -->|"blank 1"| T["Call a tool"]
  T --> M
  M -->|"blank 2"| A["Answer the user"]

Loading

Show answer

blank 1 = b (needs data), blank 2 = a (has enough). The T --> M edge is the loop-back: read the result, decide again.

Q6. Order the six agent parts the way they were assembled, given that model access is already in place:
guardrails · tools · orchestration · instructions · memory · model

• A) model, instructions, tools, memory, orchestration, guardrails
• B) model, tools, instructions, memory, guardrails, orchestration
• C) instructions, model, tools, orchestration, memory, guardrails
• D) model, instructions, memory, tools, guardrails, orchestration

Show answer

A) Model, then the role and rules, then the tools, then session memory, then the loop that drives them, then the guardrails around it.

Q7. A handler runs the same three steps every time, no judgement, no branching. On the ambition ladder it is:

• A) an agent loop, because it calls more than one tool
• B) automation, because the steps are fixed and need no model
• C) a single call plus RAG, because it may need some facts
• D) a workflow, because three steps already implies branching

Show answer

B) Fixed steps with no judgement need no model at all. Climb only when a lower rung genuinely cannot do the job.

Q8. Two candidate wirings for the same agent. Which is correct?

flowchart TD
  subgraph X["Option 1"]
    Ua["A2UI"] --> Aa["Agent: loop, guardrails, memory"]
    Aa --> Ta["Tools"]
    Aa --> Ka["Knowledge Base"]
  end
  subgraph Y["Option 2"]
    Ub["A2UI"] --> Tb["Tools"]
    Tb --> Ab["Agent: loop, guardrails, memory"]
    Ab --> Kb["Knowledge Base"]
  end

Loading

• A) Option 2
• B) both are valid
• C) Option 1
• D) neither, tools must sit above the agent

Show answer

C) The UI reaches the agent; the agent then calls tools and the knowledge base. In Option 2 the tools sit between the UI and the agent, which is wrong.

Q9. Match each gate to what must be true to pass it. Bank: a) value beats cost and the task needs an agent b) autonomy set, tools contracted, acceptance bar written c) supervised MVP runs, instrumentation on, eval suite exists d) eval passes the bar, guardrails hold, sign-off signed

1. P0 to P1
2. P1 to P2
3. P2 to P3
4. P3 to operate

Show answer

1 = a, 2 = b, 3 = c, 4 = d.

Q10. Which of these are one-way doors that need a human gate? (select all that apply)

• A) showing the passenger the alternative trains
• B) confirming a rebooking that charges a fare difference
• C) reading the PNR record
• D) issuing a refund to the passenger's card

Show answer

B and D. A charge and a refund are hard to undo. Showing options and reading a record are reversible.

Q11. Why is the P1-to-P2 acceptance bar called the hinge of the whole build?

• A) it caps the budget that P2 spending is not allowed to exceed
• B) it is the final gate before the agent reaches production
• C) it decides which model and which tools are chosen before a single line of code is written
• D) it defines good enough up front, so P3 validates against evidence, not opinion

Show answer

D) You cannot validate in P3 what you did not define in P1. The bar turns QA from an argument into an evidence check.

Q12. The brain looks up the booking, finds the cancellation cause, then fetches alternatives. For cancel RZ73KP, options? the trajectory is:

• A) lookup_pnr then get_cancellation_cause then get_alt_trains
• B) lookup_pnr then get_alt_trains
• C) get_alt_trains then lookup_pnr then get_cancellation_cause
• D) lookup_pnr then get_cancellation_cause

Show answer

A) Look up the booking, find why the train died, then fetch alternatives for the PNR and tier.

Q13. Match each loop line to its effect. Bank: a) records the model's turn in history b) binds the tool result to the call that asked for it c) ends the loop once the model stops asking for tools

1  messages.append(resp["output"]["message"])
2  "toolUseId": block["toolUseId"]
3  if resp["stopReason"] != "tool_use": return ...

Show answer

1 = a, 2 = b, 3 = c.

Q14. Why does the tool-use loop need a maximum-turn guard even before QA begins?

• A) without it the model cannot tell when it has enough information
• B) without a stop condition, a mis-stepping loop can run without end
• C) the guard is what matches the toolResult id to the toolUse id
• D) it limits how many tools the agent is allowed to register

Show answer

B) No stop condition means a runaway loop that burns tokens. The guard is also what makes did it stop an answerable question in QA.

Q15. True or False: a one-way door like a charge should sit behind a human gate, while a two-way door like showing options can run automatically.

• A) True
• B) False

Show answer

A) True. Autonomy follows reversibility: reversible actions can run on their own, irreversible ones get a human.