Cover image for Agentic workflows and vibe development – from intuition to automated collaboration

From vibe-coding to agentic workflows

In recent years, generative language models have made it possible to write code simply by describing what you want. Vibe-coding is often described as development in flow: you open a blank file, follow your intuition and let the code take shape. This way of working is fast and creative: you can rapidly prototype with ChatGPT, experiment with front-end design using Tailwind and Copilot, and build scrappy minimum-viable products with little overhead. However, vibe-coding has limitations: it scales poorly, lacks structured testing and auditing, and is not robust enough for production environments. It’s great for getting ideas moving, but taking a prototype into a reliable system requires more than flow and intuition.

Agentic workflows

Agentic AI tries to solve these limitations by making AI systems more autonomous. Vibe-coding is about making something appear to work quickly, whereas agentic AI can take abstract instructions like “build a customer database” and autonomously handle the technical details. While vibe-coding is primarily a human-to-AI interaction via natural language prompts, agentic systems extend this into an ecosystem of self-directed agents that make decisions and act with minimal supervision. These agents can perform sophisticated code reviews, recommend infrastructure optimisations and adapt to changing requirements.

Another way to frame this shift is agentic engineering: moving from coding with AI to building systems for AI agents. Agents work with clear roles and execution loops – for example a planning agent that decomposes a goal, a coding agent that writes and tests code, and a security agent that checks outputs. This makes systems more secure, scalable and auditable, but it requires a deeper design mindset focused on state, lifecycle and trust boundaries.

Standards and tools: AGENTS.md and llms.txt

If agents are to work together on complex tasks, they need structured context. Two recent initiatives are worth knowing.

AGENTS.md – a README for agents

AGENTS.md is a simple, open file standard designed to give AI coding agents a dedicated place to find context and instructions about a project. Think of it as a README for agents: a predictable place to list build and test commands, code style, PR guidelines, security notes, and other details that might clutter human-facing docs.

llms.txt – information for language models

Where AGENTS.md targets the codebase, llms.txt aims to help large language models understand websites. The proposal standardises a /llms.txt file as a single place for websites to provide models with a concise overview of content, background and links to more detailed markdown pages. The rationale is that models often have limited context windows and struggle with complex HTML structures. The file is human- and machine-readable and follows a consistent format with a title, brief description and lists of relevant markdown files. Several projects (e.g. Svelte/SvelteKit) already expose LLM-ready documentation in multiple sizes (small/medium/full) to fit different context windows.

A future with vibe-design and vibe project management

Combining vibe-coding with agentic workflows and standardised metadata opens the door to new disciplines. The same principles that make it possible to write code in flow could be applied elsewhere:

Suggestions

A practical workflow that combines LLMs and agents across the entire lifecycle could look like this:

  1. Customer acquisition Research agent builds a lead list, drafts tailored outreach, and prepares account briefs (industry, tech stack, compliance).
  2. First meetings Meeting assistant transcribes, extracts decisions and risks, and creates a concise brief. Action items are auto-synced into the workspace.
  3. Scoping & alignment LLM co-pilot drafts a one-page vision and key objectives (KRs). A governance agent flags data/privacy constraints early.
  4. PRD creation PRD agent turns the vision into a structured PRD (user stories, acceptance criteria, non-functionals). Compliance agent checks regulatory alignment.
  5. Backlog & breakdown Planning agent decomposes PRD into tickets with dependencies. Estimation agent proposes effort ranges and highlights critical path.
  6. Vibe-design Design agent proposes wireframes and flows; accessibility agent runs checks; human designers review and approve variants.
  7. Implementation Coding agents scaffold services, generate tests, and wire CI. Test agents run TDD loops; security agent performs SAST/DAST-like checks.
  8. Review & iteration Code review agent enforces style and architectural guardrails; summarisation agent compiles weekly stakeholder updates.
  9. Release & operations Ops agent manages CI/CD, canary rollout and monitors SLOs; incident agent drafts human-readable postmortems.
  10. Learning loop Retrospective agent collects metrics and lessons; knowledge agent updates AGENTS.md/llms.txt; roadmap agent suggests next bets.

This integrated flow bridges business intent, design creativity and robust technical delivery while keeping humans in decisive roles.

Conclusion

Vibe-coding lowers barriers and fosters creativity, but without structure you risk technical debt and security issues. Agentic workflows build on vibe-coding by introducing autonomous agents and standards like AGENTS.md and llms.txt, which give AI access to context and instructions. For responsible IT staff this means governance, testing and security must be prioritised. For business owners it means faster experimentation with responsible implementation. For developers the role shifts from writing code to designing systems, orchestrating agents and ensuring quality.

References