Good morning
For Leaders who want to implement AI in their organizations. Together with friends, we’re organizing the second edition of the AI_managers program, starting October 6th. So you don’t have much time to join. Join here.
For builders: Hackathon. Warsaw, October 18. Anthropic, ElevenLabs, Vercel, and more as partners. Top investors, including Credo Ventures, Dawn Capital, and more, including yours truly. Join here.
In today’s edition, among other things:
The AI Coding Trap
Steve Jobs on Hiring Great People
Where Startup AI Dollars Actually Go
The AI Healthcare Stack
Shipping with Confidence
Life Is Poker, Not Chess
Onwards!
The AI Coding Trap
We’re shipping more lines than ever — and still missing the thing customers actually feel: reliability. Demos are easy. Durable delivery is not. Code is cheap now; context, comprehension, and change-safety aren’t.
I’ve spent the last two years in code-gen AI, and code-gen Agents specifically and really loved this one:
The work isn’t “write code”; the work is “know what changed, prove it’s safe, and make it explainable.
LLMs collapse typing time. The bill shows up later, and it’s not the tokens—explaining what changed, stitching it in, and proving it didn’t break anything.
Most software effort is thinking, not typing. With AI, the “thinking” shifts after the code appears, turning reviews and fixes into the dominant cost. That’s why “10× coding” often becomes ~10% delivery in practice. Velocity moved; the bottleneck didn’t vanish.
The human experience matches the charts: fun parts go to the model; you inherit testing, deduping, docs, and deploys. The output is faster; the outcome isn’t.
This is the root of the difference between marketing copy that boasts… ‘10X faster’, and the marginal productivity gains in delivering working software… closer to 10%.
If you can’t say what changed, why it’s correct, and how you know, the speed is artificial.
There are two paths: ship fast and hope (vibe coding) or ship legibly and compound (AI-driven engineering).
Treat current agents like ultra-fast juniors: breathtaking output, brittle understanding, and no real on-the-job learning. You must provide the learning system—context, constraints, and review.
The curves diverge. Quick prototypes look magical; then complexity hits and the graph bricks. The compounding path is slower at first but survives complexity. Prototype speed is not product velocity.
Applications of sufficient simplicity can be delivered without the need for any human thinking at all… But you will hit a wall of complexity that AI is incapable of scaling alone.
The leverage is not “more code,” it’s making decisions legible across the lifecycle.
Specification first. Narrow scope, enumerate invariants and edge cases before generation. Specs are the policy the agent can read.
Documentation up front. Generate and review docs early; they become reusable guardrails and the memory agents don’t have.
Modular design. Keep contexts small so any reviewer can reason about them quickly and test them thoroughly.
Test-Driven Development. Ask the agent for tests before code; let tests guide implementation and catch regressions.
House standards via context. Feed style and patterns; auto-lint and refactor on commit.
Monitoring & introspection. Instrument logs and extract insights; make agent decisions observable and reviewable.
Value accrues in specs, tests, docs, and observability—assets that survive model swaps and team changes.
AI didn’t invent software engineering failure modes.
The fix is cultural and procedural: enforce practices that minimize rework while growing shared understanding—code reviews, incremental delivery, modular design, TDD, pair programming, quality docs, CI. It’s classical engineering, adapted for agents.
Known unknown: longer contexts and tool use might move the wall. But until agents can hold whole-system context and prove invariants, the wall moves; it doesn’t disappear. That’s why we built Shotgun.
What can you do?
Keep reading with a 7-day free trial
Subscribe to Bartek Pucek to keep reading this post and get 7 days of free access to the full post archives.