Vibe coding is old now

2026-02-05 · Source: Ben's Bites · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Advanced, extended

Summary

The article discusses the evolving landscape of AI agents, highlighting Karpathy's shift from "vibe-coding" to "agentic engineering" and the launch of a community for non-coders to build with AI agents. It covers Claude's Super Bowl campaign, which sparked controversy with OpenAI, and new updates to Droid Plugins and Claude Code, including a new `/insights` command and VS Code extension access to browsers. Mistral also released its latest Voxtral models, Voxtral Realtime and Voxtral Mini Transcribe 2, designed for live and asynchronous workloads, respectively. The piece also touches upon the challenges and future of AI agents, including security concerns like prompt injection and the general user's ability to effectively utilize these tools.

Key takeaway

For AI Architects evaluating agentic systems, understand that while current LLMs excel at bash and file system interactions, prompt injection remains an unsolved security risk. Prioritize agent harnesses that allow self-modification and custom tool creation, like Pi, over rigid MCP servers, to ensure adaptability and composability. This approach mitigates vendor lock-in and allows for tailored workflows, crucial for navigating the rapidly changing AI landscape.

Key insights

AI agents are evolving rapidly, enabling non-coders to build software and raising critical security and usability questions.

Principles

• Bash is often sufficient for effective agent tooling.
• Agentic LLMs are specifically trained for tool use.
• Self-modifying agents adapt better to workflows.

Method

Minimal coding agent harnesses like Pi use a while loop to call an LLM with four tools, allowing the LLM to execute tool calls, read/write files, and run bash commands to achieve tasks.

In practice

• Use agentic engineering for building apps without traditional coding.
• Explore Droid Plugins for bundling agent skills.
• Utilize Claude Code's /insights for workflow improvements.

Topics

• AI Agents
• Agentic Engineering
• Prompt Injection
• Claude & OpenAI Models
• AI Development Tools

Best for: AI Architect, AI Product Manager, AI Engineer, Machine Learning Engineer, Software Engineer

Related on AIssential

• Fed up with vibe coders, dev sneaks data-nuking prompt injection into their code — A developer embedded a destructive prompt injection in open-source code to deter AI agents, sparking ethical and legal debate.
• Mastering Agentic Techniques: AI Agent Customization — Effective AI agent customization combines diverse techniques, from prompt engineering to advanced RL, tailored to task needs and resources.
• Components of A Coding Agent — Agentic harnesses significantly enhance LLM coding capabilities by integrating tools, context, and memory management.
• 🙀 The 4-tool agent quietly powering OpenClaw — Simpler, user-extensible AI agents like Pi offer a sustainable alternative to complex, self-generating agent swarms.
• The AI Agent Security Surface: What Gets Exposed When You Add Tools and Memory — AI agents introduce four distinct attack surfaces requiring a new security paradigm beyond traditional LLM prompt defenses.
• No longer just a Copilot, Microsoft's AI wants to take the wheel — Microsoft's Autopilot agents promise autonomous work management, integrating deeply across enterprise applications while posing significant security and trust challenges.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Ben's Bites.