Článek
Andrew Karpathy dropped a post that hit every developer hard this week: „I’ve never felt this much behind as a programmer.“ The profession is being refactored. The bits we contribute are getting sparser. Agents, sub-agents, prompts, contexts, memory, tools, workflows — it’s a new layer of abstraction on top of everything we already know, and it feels like everyone got handed a powerful alien tool with no manual.
Meanwhile, the April 2026 numbers are brutal: 84% of developers now use AI coding tools daily. Only 29% actually trust what they ship to production.
That gap is where careers (and outages) are made in 2026.
I stopped doomscrolling the „AI is killing software engineering“ threads and did the only thing that matters: I built and tested a lightweight review system on services running real traffic. Here’s exactly how I close the trust gap without killing velocity.
The New Reality Nobody Wants to Admit
AI agents are no longer assistants — they’re autonomous workers that plan, code, test, and sometimes deploy. Claude Code Routines, Cursor Automations, multi-agent swarms… they ship features while you sleep.
But they still love writing the same expensive mistakes we’ve debugged for years:
- N+1 queries and correlated subqueries
- Cache keys that collide across tenants
- Missing composite indexes on hot paths
- No circuit breakers on external calls
- SELECT * on tables with millions of rows
The code looks correct. Tests pass. Production says otherwise — often at 3 a.m. with a pager.
This is the Developer Productivity Paradox everyone is living: massive speed gains paired with rising paranoia.
The Review System That Actually Works
I no longer treat AI output as „almost done.“ I treat it like code from a very fast, very confident junior who has never been on-call.
My enforced workflow (applied to every AI-generated change):
- Agent does its thing — generate, run tests, self-review if possible.
- 7-minute human gate — mandatory checklist before merge:
- Run EXPLAIN ANALYZE on every new or modified query. Look for sequential scans, bad join orders, or functions on indexed columns.
- Check cache key design: Does it include tenant ID + version + proper invalidation strategy?
- Verify external calls have circuit breakers, retries with backoff, and rate limits.
- Confirm rollback safety and feature flag readiness.
- Rough cost estimate: expected RPS × added latency = real money burned?
This checklist takes almost no time once you’re used to it. It caught 70%+ of the dangerous patterns the agents introduced in my tests, while still letting me keep the 3–5× velocity boost.
Get Software News’s stories in your inbox
Join Medium for free to get updates from this writer.
Subscribe
Remember me for faster sign in
The AI writes most of the lines. I own the parts that can take the service down or inflate the cloud bill.
Why Most Teams Are Still Getting It Wrong
On X and Hacker News this week, the debate rages:
- “Just vibe and ship — review is old thinking.”
- “If you can’t read the code, you can’t own the system.”
Both sides are half-right. The winning move is hybrid: aggressive AI usage + structured senior-level defense.
Teams that skip the review gate celebrate velocity metrics until the first big incident. Teams that over-review lose the speed advantage and burn out.
The sweet spot is a lightweight, consistent gate focused on production realities, not syntax.
The Exact Resources I Reach For at 2 a.m.
When the agent ships something that smells off, these three cut my investigation time dramatically:
30 Production Incidents That Cost $10K+ — Real failures from the trenches: database, deployment, cache, infrastructure, and networking. Every incident includes the exact bad code/config, detection queries/commands, the fix, estimated cost ($8K–$140K), and a prevention checklist. Many are identical to what current agents generate.
Your Database Is Bleeding Money. The Incident Playbook. — Deep PostgreSQL & MySQL guide with copy-paste ready queries for slow queries, lock contention, replication lag, VACUUM bloat, and a literal 3 a.m. decision tree when you get paged.
SQL Performance Cheatsheet — Queries That Scale — How to read EXPLAIN ANALYZE properly, design composite and covering indexes, kill the 10 anti-patterns agents adore (N+1, SELECT *, OFFSET pagination, etc.), and replace them with patterns that survive at scale. Free and always open on my second monitor.
What Changes When You Close the Trust Gap
You stop feeling 10× behind. You start feeling 10× more powerful.
The models will keep improving. The orchestration layer (agents talking to agents, tools, memory, workflows) will get more complex. But the core engineering skill — protecting production while moving fast — remains human.
Karpathy is right: the profession is being refactored. The programmers who thrive won’t be the ones who write the most code. They’ll be the ones who can direct swarms of agents and still guarantee the system doesn’t explode.
Build the review muscle now. Turn the alien tool into a force multiplier instead of a liability.
The timeline moves fast. Your on-call rotation doesn’t have to.
Source:
https://medium.com/@SoftwareNews/i-felt-10×-behind-as-a-programmer-in-april-2026-then-i-built-the-exact-review-system-that-fixes-ec9476c9ca40
