ENVISION 2026

Teaching coding in the age of coding agents.

AI with a purpose

Lucas P. Cordova, Ph.D.

Willamette University  |  October 16, 2026

// ABOUT ME

Lucas P. Cordova

Associate Professor of Computer Science
Willamette University, School of Computing & Information Sciences

I TEACH
Software developmentGame designData engineeringAlgorithmsData structuresCapstone
I RESEARCH

Software testing pedagogy and AI-infused learning tools: how feedback becomes understanding. Testing Tutor, Pedalogical, Continuity, and now SE Tutor, with undergraduate and graduate researchers at every step.

// THE PLAN

Three acts

ACT 01
The shift
Typing is cheap. Judgment is not, but judgment can be taught.
ACT 02
The lesson
Task performance is not learning. The good news: we know what is.
ACT 03
The tutor
SE Tutor: AI built to ask, not to answer.
// LIVE POLL 1

How much of the code you or your team ship today is typed by AI?

  1. Almost none
  2. Some, but less than half
  3. Most of it, and we still read every line
  4. Most of it, and nobody reads every line
  5. I do not ship code

Scan, or go to etc.ch/Ledc

QR code

ONE VOTE PER PHONE

// LIVE POLL 1 · RESULTS
// POLL 1 FOLLOW-UP

What the industry reports

80%+

of code merged at Anthropic authored by Claude (June 2026)

75%

of new code at Google AI-generated, approved by engineers (Apr 2026)

90%

of developers use AI daily; 30% trust its output little or not at all (DORA 2025)

3%

of developers highly trust AI output (Stack Overflow 2025, 49,000+ responses)

The typing is moving to agents faster than the trust is. That gap is exactly where education earns its keep.

Anthropic Institute, When AI builds itself, June 2026 · Pichai, Google Cloud Next, Apr 2026 · DORA, State of AI-assisted Software Development, Sept 2025 · Stack Overflow Developer Survey 2025

// ACT 1

The shift

Typing is cheap. Judgment is not. And judgment can be taught.

“Nobody in our team really writes code by hand anymore. I certainly don't.”

THORSTEN BALL · CO-CREATOR OF THE AMP CODING AGENT · JULY 2026

Typed, not written.
The text of source files is now cheap. Writing software was always more than that.
 

Ball, Think Harder: How I Prompt, Laracon US, 29 July 2026

The tests pass. Then what?

≈75%

of code review findings were never about bugs. They were about maintainability: can the next person understand and change this?

Mäntylä & Lassenius, TSE 2009 · Bacchelli & Bird, ICSE 2013 · Beller et al., MSR 2014 · Czerwonka et al., “Code Reviews Do Not Find Bugs,” ICSE 2015

TECHNICAL DEBT
Harder to change.
COGNITIVE DEBT
Harder to understand.
INTENT DEBT
Harder to know what it is for. Storey, ACM Queue 2026

Review moves up a level

source codeagents write it; agents check it the changespec, design decisions, behavior the intentwhat is this for? is it still right? humans move up ↑

The reviewer’s job becomes judging design and intent.

Which is exactly the set of skills we used to assume students would pick up on the job.

Adams et al., RADAR (Meta), 2026 · Paltenghi & Chandra, AfterVibe (Meta), 2026 · Brooker (AWS), 2026

// LIVE POLL 2

Now that agents type the code, which skill matters most in a new hire?

  1. Testing and knowing the code is right
  2. Design and specification
  3. Explaining code they did not write
  4. Security and defensive thinking
  5. Judgment: knowing when the AI is wrong

Scan, or go to etc.ch/2JsD

QR code

ONE VOTE PER PHONE

// LIVE POLL 2 · RESULTS
// POLL 2 FOLLOW-UP

What the people who hire said

Entry-level now means what senior used to mean.

GOOGLE + EDUCATION DEVELOPMENT CENTER, 2025
A job analysis with expert developers concluded: “a regular AI-enhanced developer of tomorrow will need to have the skills and knowledge of today’s senior developer.”
CIERA JASPAN, GOOGLE, MAY 2026
“We’re now going to need to make sure everybody is a senior engineer, including our most junior people graduating from college today.”

Kam et al., FSE Companion 2025, arXiv 2506.00202 · Seroter, Jaspan, Hammerly, Osmani, Google I/O 2026 fireside chat

// DAY ONE

Skills learned late are now day-one requirements

Specification
Testing & verification
Design
Reading code you did not write
Security
Working with non-developers

Senior courses, internships, and the first year on the job used to teach these. None of them is new, and we know how to teach every one. The question is when, not whether.

“The scars of battle help you develop a finely-tuned sense of what can go wrong. Without those detectors, modern AI tools set you up for epic failure, and the worst part is, you won’t even know you’ve failed.”

SHRIRAM KRISHNAMURTHI · ADVICE TO THE NEW COLLEGE STUDENT IN THE AI ERA · 2026

Detectors
 
 
are a person’s internal alarms. Built by struggle, which means they can be built on purpose.
// ACT 2

The lesson

Performance is not learning. The good news: we know what is.

Three studies, one lesson

−17%
Students who practiced with an unrestricted chatbot scored 17% lower on the no-AI exam than students who never had it. A tutor version that refused to hand out answers caused no drop.
Bastani et al., PNAS 2025 · about 1,000 high-school students, randomized
50 vs 67
Engineers who learned a new library with AI help scored 50 on the follow-up quiz. Those who coded it by hand scored 67. The AI group had saved two minutes.
Shen & Tamkin, Anthropic, 2026 · 52 junior engineers
74%
Of requests to a university's course AI assistant, 74% asked for the answer. Sections with the assistant ended 0.37 SD lower in final grades.
Liu et al., 2026 · 2,379 students, randomized by instructor

Answer-giving AI hurts learning. Scaffolds that withhold the answer do not.

Struggle has to land on the right thing

Desirable difficulty
A difficulty is desirable only if it exercises the skill you are building. A logic error does. A missing colon does not.
Bjork & Bjork, 2011
Productive failure
Attempt first, instruction second, then consolidate: about a third of a standard deviation across 53 studies.
Sinha & Kapur, RER 2021
The strategy gap
95% of A students probed the problem before coding. More than half of D students never did.
Denny et al., ITiCSE 2025, 1,028 students

If a course only recommends the strategies that work, the students who most need them will not adopt them. The course has to put everyone into them. That is a design problem, and design problems get solved.

Meanwhile, every tool a student touches speaks in symptoms

error  Unsanitized input on line 27
warn   Branch at line 15 not covered
warn   Null dereference at line 42
error  Complexity 23 > 10
info   14 issues, 3 blockers

Autograders, analyzers, coverage tools, and scanners say what went wrong.

Never why it matters. Never how to think differently.

So students patch the line. And a chatbot on the side just hands over the answer.

What if the tools themselves asked the questions?

// ACT 3

The tutor

AI built to ask, not to answer. SE Tutor turns tool output into inquiry, where students actually work.

Where SE Tutor comes from

2019 → 2025
Testing Tutor
NSF DUE-2013296 / 2013342 (Level 2)
Inquiry-based feedback on test suites: about 30% better test quality and concept retention, held after feedback was removed. 1,000+ students, six institutions.
SIGCSE TS 2019, 2020, 2022 (Cordova, Carver, Walia, Gershmel)
2024 → 2025
Pedalogical
AI feedback built on learning theory
Feynman Technique in the loop: 92.7 vs 74.3 against unstructured ChatGPT. Extended to security vulnerability analysis.
SIGCITE 2025 (Cordova, Mendoza, Webster, Sanders, Walia) · FIE 2025 (Sanders, Cordova, Walia)
2026
Continuity
Voice-first reasoning and reflection
Let them think out loud: a voice AI for explaining and reflecting on code in computing courses.
SIGCSE TS 2027 (Ahmed, Cordova)
2026 → 2031
SE Tutor
NSF IUSE 2540090 / 2540091 / 2540092
The same pedagogy across the whole software engineering lifecycle, in the IDE, open source, with an SDK.
Willamette · Alabama · Augusta (Cordova, Carver, Walia)

Findings are data. Pedagogy is code.

WHAT THE TOOL SAYS
semgrep: tainted input reaches SQL string
src/orders.py:27  [CWE-89]
WHAT THE STUDENT SEES
“What could a user type into this field that would change the meaning of your query? Explain in your own words why parameterization prevents that.”

The student never sees the raw message.

Tools stay dumb: they report what they found and which concept it maps to.

The SE Tutor Core decides how to teach it: cognitive load, the student’s current level, this week’s concept.

The pedagogy is what Testing Tutor validated. The new part is doing it for every tool, in the editor, on every save.

Five rungs up, three guardrails

NOTICE
Which values here come from outside your program?
EXPLAIN
In your own words, how could the query’s meaning change?
PREDICT
What would an attacker type, and what would happen?
CONNECT
Is there a test that sends malformed input here?
PREVENT
How would the design change so this cannot recur?
NEVER
quote raw tool text.
NEVER
hand over a complete solution.
ALWAYS
require the student’s own explanation before the next rung.

One tutor across the whole lifecycle

Requirements
Design
Code quality
Testing
Security
Debugging
DevOps
IN THE IDE
A VS Code extension. Intervenes on save, on test run, on commit. Sees the whole workspace, so one finding connects to another.
EDUCATORS IN CONTROL
A signed Course Profile picks plugins, starting rung, quiet hours during exams, and this week’s concept. A declarative plugin takes an afternoon.
A RESEARCH PLATFORM
Consent-gated, IRB-aligned data across three very different institutions. A/B variants per course. Open source, with an SDK.

Every sobering fact so far has an answer in SE Tutor

Performance ≠ learning
→
Answers are withheld. The student explains first.
Struggle lands anywhere
→
One concept at a time, weighted to this week’s topic.
Only A students use the strategy
→
The Course Profile puts everyone in the inquiry.
No detectors
→
Notice and predict rungs, build the internal alarm.
Review moves above code
→
Connect and prevent rungs ask about tests, design, intent.
Courses can’t change yearly
→
Tools live in swappable plugins. Pedagogy stays in the Core.
One rule underneath all six:
findings are data, pedagogy is code, and the two never mix.

Where we are, and where you come in

OCT 2026JAN 2027MAY 2027AUG 2027 Sprint 0: nowv0.1 alphav1.0 RC, spring pilotsv1.0 Foundation
PILOT A COURSE
Any level, any language the plugins support. 15 courses across three institutions in years 2 and 3.
BUILD A PLUGIN
A declarative plugin.yaml maps your tool’s findings to your course’s concepts. Templates and workshops for 50+ faculty.
JOIN THE WORKSHOP
Virtual requirements workshop, summer 2027, 20+ educators. Industry voices welcome: tell us what day one looks like.
// WHY THIS IS GOOD NEWS

The code was never the point.

For forty years the syntax tax hid the ideas. Now the typing is cheap, and what is left is what we always wanted to teach: specification, verification, design, judgment, and the ability to say why.

Struggle still works when it lands on the right thing. Tools can now put it there, for every student, every time they save.

Our students get to become engineers earlier than we did.

THANK YOU

Questions?

Lucas P. Cordova, Ph.D.

School of Computing & Information Sciences, Willamette University

SE Tutor · NSF IUSE 2540090 / 2540091 / 2540092
University of Alabama (Jeffrey Carver) · Augusta University (Gursimran Walia)

References

  1. Ball, T. Think Harder: How I Prompt. Laracon US, July 2026. · Anthropic Institute, When AI builds itself, June 2026. · Pichai, S. Google Cloud Next, Apr 2026.
  2. DORA, State of AI-assisted Software Development, Sept 2025. · Stack Overflow Developer Survey 2025.
  3. Mäntylä & Lassenius, TSE 2009 · Bacchelli & Bird, ICSE 2013 · Beller et al., MSR 2014 · Czerwonka, Greiler, Tilford, ICSE SEIP 2015.
  4. Storey, M.-A. From Technical Debt to Cognitive and Intent Debt. ACM Queue 24(2), 2026; arXiv 2603.22106.
  5. Adams et al. Automating Low-Risk Code Review at Meta: RADAR. arXiv 2605.30208, 2026. · Paltenghi & Chandra, AfterVibe. arXiv 2607.09900, 2026. · Brooker, M. Thread on X, Sept 2026.
  6. Kam et al. What Do Professional Software Developers Need To Know To Succeed in an Age of AI? FSE Companion 2025; arXiv 2506.00202. · Seroter, Jaspan, Hammerly, Osmani. Google I/O 2026 fireside chat.
  7. Krishnamurthi, S. Advice to the New College Student in the AI Era. 2026.
  8. Bastani et al. PNAS 122(26), 2025. · Shen & Tamkin, How AI Impacts Skill Formation, arXiv 2601.20245, 2026. · Liu et al. EdWorkingPaper 26-1598, 2026 (preprint).
  9. Bjork & Bjork, 2011 · Sinha & Kapur, RER 2021 · Denny et al., ITiCSE 2025 (arXiv 2504.11723).
  10. Cordova & Walia, SIGCSE TS 2019 · Cordova, Carver, Walia, Gershmel, SIGCSE TS 2020 · Cordova, Carver, Walia, SIGCSE TS 2022 (Testing Tutor).
  11. Cordova, Mendoza, Webster, Sanders, Walia. Not All Chatbots Teach. SIGCITE 2025. · Sanders, Cordova, Walia. FIE 2025 (Pedalogical for security vulnerability analysis). · Ahmed & Cordova. Let Them Think Out Loud. SIGCSE TS 2027.
  12. Several industry and learning-science sources above were first collected in Bill Pugh’s talk The CS Curriculum in the Age of Coding Agents (UC San Diego, Oct 12, 2026), whose full reference list is at billpugh.github.io/talks/2026-10-12-UCSD.