Atlanta DevCon 2026

Where AI belongs in your first five years, and where it doesn't.

The job is moving from writing code to owning what ships. Here's the talk on one page.

Understand where AI is taking the software engineering career.
Decide how to use it in light of that.
Learn what makes a great engineer.

Where AI belongs.

No stage of building software goes away. In each one, AI takes on more of the work, and you own the rest.

Stage	AI increasingly does	You own
Requirements	Draft the doc from meeting notes	Decide what's worth building
Architecture	Sketch first-draft options	Keep the design coherent across fifty changes
Implementation	Type boilerplate	Know what the code should do
Testing	Write routine unit tests	Decide what proves it works
Code review	Nitpick style	Judge whether it's right
Deployment	Run releases	Own what ships and who it affects
Operations	Watch dashboards for routine alerts	Debug a live system under pressure

SHIP.

What made junior engineers succeed before AI still decides who thrives now. Scrappy means resourceful, not reckless.

Scrappy

Before AIFound answers on their own. Read the source. Built a quick prototype to learn.

With AIUses AI to get unstuck fast. Understands the answer before moving on. Resourceful, not reckless.

Humble

Before AISaid 'I don't know.' Owned mistakes. Asked for review.

With AISaid 'I don't know.' Owned mistakes. Asked for review. Use AI to learn, not to cover mistakes or seem smart.

Inquisitive

Before AIAsked why, not just what. Read other people's code.

With AIAsks AI to explain. Then checks the explanation. Still ask others 'why.' Learn from senior engineers and practitioners.

Passionate

Before AICare about how it works and who it's for. Solve problems. Enjoy your work.

With AICare about how it works and who it's for. Solve problems. Enjoy your work. Don't ship AI slop.

If you can't slow down.

Before every commit, ask: if this breaks in production, could I debug it right now without AI? Write down anything you shipped but don't understand, and review it weekly. This is triage. Real learning time still wins, so protect some every week.

If you lead engineers

Measure juniors on more than PRs merged. Give new hires reps without AI. Incentivize learning, failing, and questions over speed.

The evidence.

Every number from the talk, with its source. Independent means academic, government or nonprofit. Vendor means a company that sells something related. Read the caveats.

Software developers: projected +10% from 2025 to 2035, about 174,700 jobs.
U.S. Bureau of Labor Statistics, Occupational Outlook HandbookindependentProjections, not outcomes. BLS says its AI exposure categories do not measure AI effects on employment.
Computer programmers: projected to decline 7% from 2025 to 2035.
U.S. Bureau of Labor Statistics, Occupational Outlook HandbookindependentProjections, not outcomes.
Workers aged 22 to 25 in highly AI-exposed jobs: employment about 19% below where it would be had it kept pace with less-exposed peers. Experienced workers show no comparable gap.
Stanford Digital Economy Lab, "Canaries in the Coal Mine?" (Aug 2026 revision)independentDescriptive, not causal. Some differences began before generative AI.
4.5% of Q1 2026 software development postings were entry-level; 69.3% were senior.
Indeed Hiring LabvendorIndeed postings only. Seniority assigned by an algorithmic classifier.
Young workers (20 to 24) slipped from 16.4% to 15.5% of employment in highly AI-exposed occupations, mostly because fewer moved straight into jobs, not because of layoffs.
Federal Reserve Bank of Dallas, Jan 2026independentMay not be causal.
The length of software tasks AI can finish half the time has doubled about every 7 months since 2019, and about every 4 months since 2023.
METR, Time Horizon 1.1 (Jan 2026)independentWide error bars. Sensitive to which tasks are in the suite.
On SWE-bench Verified, top AI performance rose from 60% to near 100% in a single year.
Stanford AI Index 2026independentOne benchmark. Benchmarks are cleaner than real work.
Doubling the time horizon does not double the degree of automation.
METR, limitations of time horizons (Jan 2026)independentMETR's own caution about reading its trend.
About half of AI patches that passed the automated tests would not have been merged by the project maintainers.
METR, March 2026independent4 maintainers, 296 patches, 3 repositories, mid-2024 to late-2025 models, no chance to iterate. Maintainers re-approved 68% of the human reference patches.
In one study, a coding agent predicted 73% success on its own work and achieved 35%.
Kaddour et al., arXiv 2602.06948independentOne model on 100 SWE-bench Pro tasks. Every model tested was overconfident.
In one agent setup, 57% of solutions worked and 11.8% were secure.
SusVibes, arXiv 2512.03262 (ICML 2026)independentOne of 12 agent settings tested, 186 tasks.
Developers learning a new library with AI scored 17 points lower on a later quiz (50% vs 67%). Those who asked AI to explain did far better than those who delegated.
Anthropic, Shen and Tamkin, arXiv 2601.20245vendorn=52, one library, one quiz. Not a measure of long-term deskilling.
After adopting an AI coding agent, repositories saw a large, short-lived speed gain and a lasting rise in code complexity (about +41%) and static-analysis warnings (about +30%).
He et al., Carnegie Mellon, arXiv 2511.04427independent806 repositories, difference-in-differences.
On underspecified tasks, AI agents rarely ask clarifying questions unless prompted to.
Ambig-SWE, ICLR 2026independentBenchmark setting.
When agents extend their own code over many steps, quality erodes. No agent solved any problem end to end.
SlopCodeBench, arXiv 2603.24755independentBenchmark. Best checkpoint solve rate about 17%.

Stay in touch.

I talk about AI at work and early careers.