Vibe coding is dead. Here's what replaced it
What I actually do now instead of writing code.
A year ago I wrote 80% of my code and used AI for the rest. That ratio flipped a few months ago.
Now I decompose problems, assign agents, and review their output. I don't think "writing code" describes what I do most days.
Karpathy coined "vibe coding." Then he killed it. His February take: you're not writing code 99% of the time. You're orchestrating agents.
Why It Matters
Most teams I work with are still in chat-with-AI mode. Prompting one model, one conversation at a time using the browser.
Meanwhile, four major model launches in February all led with multi-agent orchestration as the headline capability, not better benchmarks or bigger parameter counts. Claude's agent teams built a 100,000-line C compiler across 2,000 coordinated sessions. Kimi K2.5 runs 100 sub-agents making 1,500 tool calls per task.
The capability is here and the gap is workflow rather than tools.
The Part Most People Miss
SonarSource surveyed 1,100+ developers globally in January 2026. AI now accounts for 42% of all committed code. But 96% of developers don't fully trust it — and only 48% always verify it before committing. The review burden is real, and most teams aren't even doing it.
That's the whole lesson of vibe coding. It optimized for generation speed and ignored everything downstream.
Agentic engineering flips it: you plan and design the system first, define the boundaries and contracts, then let agents execute inside constraints you built. The teams pulling ahead aren't using better models. They're treating agent orchestration like distributed systems engineering. Same decomposition, same contracts between components, same observability.
Lessons Learned
- "AI is a power tool, not a replacement." — Engineers who use AI ship faster. Engineers who only use AI ship garbage. The skill is knowing when to prompt and when to think.
- "Plan first, ship fast, iterate features." — AI moved the bottleneck. Writing code is no longer the slow part. Thinking about what to build, how it fits together, what breaks at scale — that's where the time goes.
- "Document your decisions, not just your code." — LLMs don't store context. If you want an AI assistant to move fast on an existing codebase, it needs documented decisions to load.
Worth Reading
- Agentic AI Strategy— Argues orgs are failing at agentic AI because they're automating existing processes instead of redesigning for agent-native workflows.
- Designing Effective Multi-Agent Architectures — Practitioner guide for moving from single-agent to production multi-agent systems — covers patterns that actually hold up.
- Testing AI Coding Agents: Cursor vs. Claude, OpenAI, and Gemini — Real benchmark with real methodology. Shows what works when you move past the marketing and actually test these tools.
If your team is trying to make agentic workflows stick and it's not clicking, I consult on exactly this — helping teams move from chat-based AI to production agentic workflows. Hit reply and tell me what you're working with.
— Collin
