Three agents, one deployed website: what an agentic pipeline actually looks like

"Agentic" is one of those words that means everything and therefore nothing. The fastest way to make it concrete is to stop talking about agents in the abstract and give each one a job, an input, and a handoff. When you do that, an agentic system stops looking mysterious and starts looking like a well-run assembly line that happens to be made of models.

DocsLM is a good example, because the job is easy to state: upload up to nine PDFs, and get back a deployed website with an embedded chatbot that answers questions from those documents, live at a public URL in minutes. Three agents do the work, and each has a role you could put on a business card.

The Architect

The Architect Agent runs an agentic RAG loop over the uploaded documents, querying them repeatedly until it has extracted a complete site structure. This is the reasoning-heavy step: not "summarize these PDFs" but "keep interrogating them until you understand them well enough to lay out a site." Its output is a structure, not prose, and that structure is the handoff.

The Designer

The Designer Agent takes that structure and generates the actual HTML and CSS, embedding the chatbot widget into the page. It does not need to understand the documents; the Architect already did. It needs to turn a known structure into a real interface. Narrow job, clean input, and it is far more reliable for the narrowness.

The Deployment agent

The Deployment Agent pushes the generated site to Vercel's API and returns a live public URL. It is the least glamorous of the three and arguably the most important, because it is the step that turns a generated artifact into something a person can actually open and share.

Roles beat one giant prompt

Around the agents, Django REST handles uploads and webhook communication and n8n orchestrates the multi-step workflow, keeping the UI concerns and the AI orchestration as clean separated layers. But the design lesson is the split itself. One mega-prompt asked to "read these PDFs and build and deploy a site" fails in ways nobody can debug. Three agents with defined jobs and explicit handoffs fail in ways you can point at.

An agent is not a magic worker. It is a job plus a handoff. Get those boundaries right and the architecture, not the model, becomes the thing that makes the product work.