Building the Right Things Well with Agents: Test driving and learning loops to Close Accountability Gaps in AI-Assisted Development

Agile 2026 · Emerging Technologies · Maryland

Ken Judy · Senior Partner, Stride Consulting
kenjudy.us · stride.build

Agenda

01The Problem

02Why Review Fails

03The Framework

04Live Demo

05Measurement

06Discussion

Today’s Session

Using a Plan-Do-Check-Act framework with XP disciplines to close the accountability gap unstructured adoption leaves open.

Evidence, framework, live demo, and discussion. 105 minutes.

Career Arc: Org Seniority vs.
Hands-on Coding (1997–2026)

← Org Seniority / Hands-on Technical Depth →
(1=lowest, 5=highest)

Org. Seniority Hands-on Technical Depth

1,718 contributions in the last year

Contribution settings ▼

JulAug SepOct NovDec JanFeb MarApr MayJun

Mon

Wed

Fri

@stride-nyc

Activity overview

Contributed to stride-nyc/litrealms-poc,
kenjudy/pdca-agentic-coding-f…,
stride-nyc/jira-beacon-team-da…
and 29 other repositories

86% Commits

10% Pull requests

4% Issues

01

The Problem

The tools are working. The outcomes are not.

AI Coding Is Now Standard Practice

90%

use AI at work

DevOps Research and Assessment 2025
~5,000 technology professionals
Conducted by Google Cloud

80%+

report productivity gains

DORA 2025

61%

never use agent mode

DORA 2025

Chat and predictive text completion are where developers actually spend their time with AI.

Adoption is not driving return on investment

↑59% report code quality improvements – DORA 2025

↑70% report at least some confidence in generated code quality – DORA 2025

↓Every 25% increase in AI adoption leads to a 7.2% decrease in delivery stability – DORA 2024

↓10x increase in code duplication over 3 yrs – GitClear 2025

↔Negligible increase in delivery throughput – DORA 2025

“

The value of AI is not going to be unlocked by the technology itself, but by reimagining the system of work it inhabits.

DORA 2025

AI Adoption Is Degrading Code Quality

10x

duplicated code blocks 2022–2024

GitClear 2025
211M lines of code

18%

of inconsistent clone changes contain faults, varying by system type

Juergens et al.

18.42%

of buggy code clones propagate to copies

Mondal et al.

“Growing Evidence that AI-Generated Code Optimizes for the Short-Term.” — GitClear

02

Why Code Review Fails

The accountability gap and what research shows code review is actually good at.

Code Review Was Never Good at Finding Bugs

What We Expect

Find defects

Catch correctness and security issues

What It Delivers

Understanding the change

Code improvements and readability

Team norms and knowledge transfer

Bacchelli & Bird, ICSE 2013 · Sadowski et al., ICSE-SEIP 2018

“

Finding defects was the top motivation for 44% of developers while only 14% of review comments address defects and most of those are small logical low-level issues

Bacchelli & Bird, ICSE 2013

AI Makes This Worse at Every Level

Speed: Controlled studies show developers complete tasks up to 55.8% faster with AI assistance. (Peng et al. 2023)

Size: AI-generated code arrives in larger batches with a 10x increase in duplicate code blocks between 2022 and 2024. (GitClear 2025)

Confidence: 39% of developers report little or no trust in AI-generated code quality. (DORA 2024)

03

PDCA

Test driving and learning loops disciplines in Plan-Do-Check-Act at the human-AI interaction level

Software is a design activity.

Manufacturing is repetition at scale.

Design is discovering the right solution under uncertainty.

(thank you Mary Poppendieck — author of Lean Software Development)

When speed is applied to undisciplined design, it increases up output without improving judgment. You get more of the wrong things faster.

Set based design develops alternatives within known constraints and validates with customers.

1.User centered focus

2.Non-functional constraints

3.Prototype within constraints

4.Validate

Plan

Act

Do

Check

Plan what you are doing

Act on anything that went wrong to avoid errors of the same nature in future

Do what you said you’d do

Check that you did it right

Developed by Walter A. Shewhart in the 1920’s.
Popularized as part of the Toyota Production System.

Structured Prompting Works

61%

Defect reduction from Plan-Do-Check-Act cycle in software development (in a case study)

Ning et al., 2010

1–74%

Improved performance of structured prompting vs. ad-hoc, depending on technique and task complexity

Sahoo et al., 2024

“

The recommended workflow has four phases. Explore, Plan, Implement, Commit.

Best Practices for Claude Code (Anthropic)

The PDCA Cycle Applied to AI Coding

P

PLAN

Analyze codebase. Identify existing patterns and constraints (Breadth-wise).

Plan the implementation before writing any code (Depth-wise).

Developer reviews and approves before proceeding.

D

DO

Test-driven implementation in small, atomic increments. Testing has anti-patterns context.

Enforces “called shots”, red-green discipline, architectural safety, and git history preservation.

C

CHECK

Completion analysis: agent reviews session transcript and generated code against original plan. Explicit definition of done beyond functional testing.

Includes a check for refactoring opportunities.

A

ACT

Micro-retrospective. Agent analyzes what worked and suggests targeted refinements to prompts and interaction patterns for the next cycle.

Working Agreements: Human Accountability in Writing

Commitments made by the developer or collectively by a team

Guidelines for when to intervene

Asserts accountability for code the developer did not write

Refined through retrospection

Introducing the DORA AI Capabilities Model

https://cloud.google.com/blog/products/ai-machine-learning/introducing-doras-inaugural-ai-capabilities-model

PDCA Mapped to the DORA AI Capabilities Model

Plan
Do

User-centric Focus

→

Team Performance

Strong Version Control Practices

→

Code Quality

AI-accessible Internal Data

→

Individual Effectiveness

Working
Agreements

Working in Small Batches

→

Product Performance

Clear and Communicated AI Stance

→

Friction

Quality Internal Platform

→

Throughput

Check
Act

Healthy Data Ecosystems

→

Organizational Performance

PDCA as a Claude Skill/Prompts

The framework’s Claude skill/prompt templates and supporting tools are publicly available.

PDCA Framework Prompts A disciplined framework for AI-assisted code generation github.com/kenjudy/pdca-code-generation-process

Code Quality Metrics Script and Github Actions for measuring AI Code Drift github.com/stride-nyc/code-quality-metrics

Steve Yegge’s Beads A memory upgrade for your coding agent github.com/gastownhall/beads

04

Live Demo

Using Claude Code with author’s PDCA skill

Running the demo

You can run the demo yourself.

Demo Instructions Setup the necessary requirements and step through the demo github.com/kenjudy/pdca-code-generation-process/blob/main/presentations/XP%202026/pdca-demo-instructions.md

Presentation Video, Demo Video TBD I will record the presentation including the demo and the demo in isolation.

05

Measurement

Detect drift before the debt compounds.

“

When a measure becomes a target, it ceases to be a good measure.

Goodhart’s Law

Measuring AI Adoption Alone Is Dangerous

What most teams measure

Lines of AI-generated code accepted

Pull requests merged per developer

Velocity increase

Tasks completed

What goes unmeasured

Commit size and test discipline trends

Code duplication and churn rates

Delivery stability

Actual value delivered

The Same Rate of AI Adoption Can Have Different Consequences

Foundational Challenges

low throughput, instability, burnout, and weak product performance

High adoption = accelerating debt.

Output metrics rise while delivery stability falls. AI makes the gap between what teams produce and what they can sustain wider, faster.

High Impact, Low Cadence

Strong outcomes and individual effectiveness, but low delivery frequency and high instability

High adoption = growing invisible risk.

Product metrics look healthy. Delivery instability is building underneath. Adoption metrics will never show it.

Harmonious High-Achievers

High throughput and stability together, low burnout.

High adoption = compounding gains.

Adoption and delivery metrics move together because the strong practices are in place to connect them.

DORA Seven Team Archetypes

Measuring Code, Platform, and Organizational Healths

Measurable from GitHub commit history

Commit size and sprawl

Test-first discipline rate

Velocity trend

Commit message quality

github.com/kenjudy/code-quality-metrics

Requires additional tooling or surveys

Delivery stability and throughput (CI/CD data: DX, LinearB)

Code duplication and cloning (AST analysis: GitClear)

AI policy clarity and platform quality (internal audit)

Developer well-being and burnout (team survey)

06

Discussion

Your team. Your archetype. Your intervention priority.

Human-centric agile disciplines are not constraints on AI capability. They are a mechanism by which AI capability is converted into delivery value.

What would it take for you to commit to attempting a structured practice before your next AI coding session?

Resources

The PDCA skill and measurement tools are publicly available under CC 4.

PDCA Framework Prompts github.com/kenjudy/pdca-code-generation-process

GitHub Actions / AI Drift Measurement Tool github.com/stride-nyc/code-quality-metrics

InfoQ Article infoq.com/articles/PDCA-AI-code-generation

Agile Alliance Blog Post agilealliance.org/reducing-ai-code-debt

Contact Me

Please reach out to learn more or to share your own approaches to agentic coding.

LinkedIn www.linkedin.com/in/kenjudy/

Github github.com/kenjudy

Personal Website kenjudy.us

Employer Website stride.build

I used Claude for brainstorming, argument review, and drafting assistance across multiple revisions. I personally verified all sources and made all final content decisions. I take full responsibility for the accuracy, originality, and quality of this work. [References]