From syntax to orchestration.
Bartłomiej Krupa
Agentic Engineering · LLM Optimization · AI Cost Reduction
Models now turn plain language into working software. Writing every line stopped being the job. Directing the agents that ship it is.
No hype. Only what ships.
What_Is_Changing
Prompts replaced source code
By late 2024, AI was already writing more than a quarter of new code at Google. Today the work is writing prompts. Rarely just one. Usually several, spread across different models, synthesized by hand until the result holds up.
The skill moved up the stack
Agent-written code already beats what most programmers produce by hand. Framing the problem, picking the right tools, and controlling what context the model gets to see: that's the work left for engineers now.
Working is cheap. Dependable is not.
Anyone can describe an app in plain English and have something running by evening - the floor has never been lower. Secure, maintainable, affordable at scale? That bar hasn't moved. The gap between the two is where engineering lives now.
Principles
Never ship what you can't explain
The work can be delegated; the understanding can't. Agents run around the clock and respond when production breaks, but the engineer still decides what gets touched and when. What you can't explain, you can't fix, trust, or own.
Taste doesn't automate
Code gets generated faster than anyone can read it now. Deciding what's worth building, keeping it simple, catching the answer that looks right but is quietly wrong - that's the part still on you.
A minimal Claude Code harness. One command.
If you vibe-code with Claude Code, this is the harness to start from: a lean CLAUDE.md skeleton, a verify-before-done loop, three read-only subagents, and credential deny rules. On four adversarial prompts it scored 4/4 where a 129-instruction CLAUDE.md scored 3/4 — [measured, not a benchmark](/articles/claude-md-adherence-bench). Every file maps to a rule published on this site.
$npx leanharness
10 files · 0 dependencies · nothing overwritten
I'm Bartłomiej Krupa. I write about harness engineering, agentic systems, and keeping LLMs cheap in production.