Agentview: CMS for Agents

Learn how building a CMS‑style scaffolding—conversation viewer, collaboration tools, versioning, and integrations—streamlines conversational‑agent development, with a demo of the open‑source agentview project.

Overview

In recent months I’ve built a bunch of conversational agents for ecommerce (one for famous 1B$+ brand). It’d seem that the most important part of such job should be AI engineering: business goals analysis, crafting prompts, tools and AI pipeline, evals, error analysis etc.

The hard lesson: it’s not. There is something else.

I think I spent more time on building scaffolding around my agent than the agent itself:

conversation viewer for team members, with custom business-specific views (look at your data!)
collaboration - domain experts and business stakeholders should be able to test, comment, feedback and discuss agent outputs (like in Google Docs or Notion)
storage, stateful API, deployment
channel integrations (connect input like email/whatsapp, or output: zendesk/intercom for human escalation)
search
version management
management of users & teams
client requested being able to play with the prompts and knowledge base, never managed to get to that point

This scaffolding feels almost like a “CMS”, but for agents. And I think we might need one.

In this presentation:

I’ll share my experiences what scaffolding should look like so that clients love the process
I’m gonna show an experimental open-source project agentview, which goal is to dramatically reduce the time to build one (inspired by architecture from Sanity CMS, Shopify and … wordpress ;)).

Tech stack