From Single-Turn to Multi-Step Execution
Differentiate Step DoD from Overall Completion Criteria, preventing the systemic failure trap where every step passes but the task outcome fails.
From Single-Turn to Multi-Step Execution
In Module 1 Lesson 6, you learned Definition of Done (DoD) — the explicit contract describing what a "good response looks like" for a single output. When AI shifts from single-turn Q&A to multi-step execution workflows, the product management problem changes completely: it is no longer "1 DoD for 1 output", but a sequential chain of actions where each link requires its own acceptance criteria to ensure the overall job is delivered properly.
Running example: RefundBot — an agent automating refund requests for e-commerce order #4821.
1. Distinguishing Step, Step DoD, and Overall Completion Criteria
To govern an agentic workflow across multiple steps, PMs must differentiate 3 core concepts:
- Step: An independent atomic unit of action performed by the agent (invoking a tool, computing a formula, querying a database, calling a third-party API). It is a concrete job action rather than a chat response.
- Step DoD (Step Definition of Done): The verification criteria ensuring that specific step executed correctly from a technical and local data standpoint. Example: "Refund API returned HTTP 200" or "Extracted leave days count from user prompt."
- Overall Completion Criteria: Product-level acceptance criteria verifying that the entire workflow achieved the user's ultimate business intent, rather than merely verifying that "all intermediate steps completed execution."
Step DoD vs. Overall Completion Criteria
Click each step to see how individual Step DoDs can PASS while the entire workflow still fails.
3. Calculate Refund
PASS (Technical) / FAIL (Business)Step Definition of Done (Step DoD):
Outputs a valid numeric currency amount (e.g. $50).
⚠️ TRAP: Agent forgot to deduct $5 shipping fee. Step DoD only verified 'returned a number', so it passed despite wrong math!
Customer receives the exact refund amount per policy (shipping deducted), backed by a gateway transaction ID and matching receipt.
Passing local technical Step DoDs does not guarantee the business outcome without holistic outcome verification.
2. Concrete Example: When Every Step Passes DoD but the Task Fails
The most common trap for PMs newly working with agents is assuming that: Passing all Step DoDs guarantees achieving the Overall Completion Criteria.
Consider RefundBot's 5-step flow for "refund order #4821":
- Verify Order: Check that the order exists and is within the return window → Step DoD: Valid order state returned.
- Check Policy: Confirm no prior refund recorded and valid defect reason specified → Step DoD: Passes policy filters.
- Calculate Refund Amount: Deduct shipping fees and applied voucher values → Step DoD: Returns a valid numeric value.
- Call Payment Gateway: Execute payout API to customer's card → Step DoD: Gateway returns HTTP 200.
- Send Notification: Dispatch email receipt to customer → Step DoD: Email delivery status = sent.
The Failure Mode: If in step 3, the agent miscalculates the formula (forgets to deduct the $5 shipping fee) but still outputs a valid number $50, step 3's Step DoD Passes. Steps 4 and 5 execute smoothly and also pass their respective Step DoDs. However, the customer receives an overpayment, violating financial policy.
Overall Completion Criteria catches this systemic flaw: "The customer receives the exact refund amount per company policy, supported by a valid payment gateway transaction ID and a receipt email reflecting the exact matching amount." This verifies the actual business outcome, not just execution mechanics.
3. Comparison: Step DoD vs. Overall Completion Criteria
| Dimension | Step DoD (Step Level) | Overall Completion Criteria (Workflow Level) |
|---|---|---|
| Scope | Locally verifies an isolated atomic action | Holistically verifies end-to-end ecosystem state |
| Nature | Technical / Execution status (HTTP 200, non-empty) | Business / Real value delivered to the user |
| Hidden Risk | Every step passes while compounding logic/data fails | Cannot isolate mid-trajectory crashes without Step DoDs |
| Defined By | Tech Lead / Developer collaborating with PM | Owned entirely by the PM |
4. Analogy: Junior Employee Checklist vs. Expert Consultation
This operational shift is analogous to delegating a SOP checklist to a junior employee versus asking an expert for advice:
- Expert Consultation (Single-turn): You ask a clear question and receive a self-contained, validated answer.
- Junior Employee Workflow (Multi-step Agent): You must supply a checklist with acceptance gates at each milestone (Step DoD), plus a final sign-off criteria (Overall Completion Criteria). An employee can check every box on their clipboard while still delivering a defective output if the checklist lacks holistic outcome reconciliation.
Exercise 38.1: A 200-person enterprise wants to launch LeaveBot on Slack — employees simply type /leave 3 days starting Sep 2nd instead of navigating legacy HR portals.
Company Policy:
- Leave ≤ 2 days: Agent auto-approves if employee has sufficient balance.
- Leave ≥ 3 days: Agent routes request to Direct Manager on Slack for approval before updating HR records.
- All cases: Upon completion, Agent updates the attendance database and sends a confirmation message to the employee.
Your Task:
- Break down the workflow into concrete steps (branching for ≤ 2 days and ≥ 3 days).
- Author Step DoDs for each individual step.
- Author the Overall Completion Criteria for the entire workflow to catch silent semantic failures where steps pass technically but business logic fails.