Master AI Agents with Three Engineering Frameworks: WBS, Delivery Contracts, Scrum

Core Thesis: AI Agent Success Is a Management Problem

After 100 days of embedding WorkBuddy agents into real engineering activities—architecture analysis, test case authoring, standardized document generation—the author concludes that the bottleneck is never model capability. The limiting factor is the human operator's ability to manage the agent. Three specific management actions are missing: task management, delivery management, and iteration management. Each maps to a decades-old, battle-tested engineering framework.

Three Missing Management Actions and Their Mature Counterparts

Task management missing → PMP Work Breakdown Structure (WBS) for deliverable-oriented decomposition and human-AI division of labor.

Delivery management missing → TOGAF-inspired delivery agreements that fix scope, process, steps, constraints, artifact format, and acceptance criteria.

Iteration management missing → Scrum sprint cycles (Product Backlog, fixed sprints with 1–2 rule changes, retrospective with fixation or rollback) for continuous skill optimization.

Tool 1: WBS Work Package Decomposition for Human-AI Division

Deliverable-Oriented Decomposition

WBS's core principle: decompose toward tangible deliverables, not actions. A valid work package must meet three criteria:

Has a clear, concrete deliverable.

Boundaries are explicit—what is included and what is not.

Can be independently verified without depending on other tasks.

Granularity rule: decompose until each package is independently verifiable; stop there. Over-decomposition wastes verification time; under-decomposition pushes errors to the end.

Concrete Example: Test Case Development

The author breaks test case creation into three sequential deliverables, each with its own verification checkpoint:

Parameter mapping verification table

Individual test cases

Test case difference comparison list

Each deliverable is produced, then immediately checked. Contrast this with action-based decomposition (“read requirements → write cases → organize”), which yields no intermediate artifacts—errors surface only at the end.

Human-AI Division of Labor

Principle: fixed-rule, verifiable, repetitive work goes to the agent; judgment, trade-offs, risk assessment, and standard-setting stay with the human.

Human responsibilities (high-value, non-automatable): define goals and test coverage scope; make design trade-offs and risk judgments; perform risk assessment and final text sign-off; set standards and rules.

Agent responsibilities (standardized, verifiable, repetitive): retrieve and organize interface documentation; draft base test cases and fill steps; unify formats and cross-check fields; run checklist verifications and repetitive comparisons.

Tool 2: Delivery Agreements to Fix Business and Acceptance Rules (TOGAF-Inspired)

Even with perfect task breakdown, output drifts without explicit business rules and acceptance standards. The author adapts TOGAF's architecture contract thinking into a six-dimension delivery agreement that tells the agent what it may do, what it must not do, and what “done” looks like.

Six Dimensions of a Delivery Agreement

Scope – what is in/out of bounds. Omission consequence: agent silently expands scope.

Process – execution sequence logic. Omission consequence: agent follows its own habit.

Steps – input and output of each step. Omission consequence: no intermediate checkpoints.

Constraints – forbidden modifications (baseline parameters, naming conventions, etc.). Omission consequence: agent “helpfully optimizes” them away.

Artifact format – structure, layout, document standards. Omission consequence: deliverable shape is random.

Acceptance criteria – itemizable pass/fail standards. Omission consequence: only manual eyeballing works, never finishes.

Acceptance Criteria: The Highest-Value Dimension

The author stresses that acceptance criteria are both the most valuable and the most frequently omitted. “Compliant with specification” is not a criterion because it cannot be judged. A real criterion is: “Parameter table covers all 12 signals, each with value range and default value”—every item can be ticked off.

Practical technique: write acceptance criteria as a checkable table. Example:

Branch scenarios ≥ 8

Boundary values: at least 2 per parameter

Regression items: mapped to previous version with retain/delete annotation

Each line is a binary check; unchecked means incomplete.

Tool 3: Scrum Sprint Iteration for Continuous Skill Optimization

Common Pitfall: Batch Rule Changes

Most users modify multiple skill rules at once. With several variables changing simultaneously, causality is lost—improvements cannot be attributed, regressions cannot be isolated, and no reusable experience accumulates.

Scrum Adaptation for Skill Evolution

Product Backlog (rule todo list): every optimization idea, bug fix, or rule tweak enters the backlog first—no ad-hoc changes.

Fixed sprints, small steps: each sprint changes only 1–2 rules. This is engineering experiment discipline: single-variable changes enable precise effect measurement. Sprint cadence (e.g., weekly or daily) is fixed.

Retrospective, fix or rollback: after each sprint, compare before/after output. Effective rules are promoted into the skill library; ineffective ones are reverted to the prior version.

Skills are not written once; they grow through repeated, controlled sprints—becoming organizational assets.

Closed-Loop System: Sequential Dependency of the Three Tools

The three frameworks are not parallel tips; they form a strict sequence covering pre-execution, during-execution, and post-execution:

WBS (pre): decompose work, define human-AI boundaries.

Delivery Agreement (during): lock down standards for each decomposed package.

Scrum Sprint (post): iteratively refine the skill rules that drive the packages.

Skipping any link breaks the loop. Only the full chain yields engineering-grade, reproducible, team-shareable human-AI collaboration.

Conclusion: Competitive Edge Is Management Mastery

AI tools are commoditized; models are equally accessible; technical parity is near-total. The differentiator is the ability to apply mature engineering management systems—WBS, delivery contracts, Scrum—to govern agents. Professionals who master this integrated management stack turn AI into stable, reliable, continuously evolving enterprise assets. The future of human-AI competition is not who prompts better, but who builds better AI management systems.

References

Primary practice: author's WorkBuddy 100-day engineering deployment (July–October 2026).

WBS theory: PMBOK Project Management Body of Knowledge.

Delivery agreement thinking: TOGAF implementation governance and architecture contracts.

Sprint/Product Backlog theory: Scrum Guide.

All frameworks are adapted for engineering contexts, not literal translations.