Step Zero: Centralize Tools in MCP Servers

▶ Watch (0:44)

Deco CMS runs storefronts for high-volume Brazilian retailers. Their customers use VTEX, an e-commerce platform similar to Shopify. Rodrigues’s team automatically turned VTEX’s entire API into an MCP server with tools. They connected other systems too: ClickHouse, GitHub, and HyperDX for error logs. This gateway gives agents a single place to call the tool they need.

Step One: Teach Agents Your Domain Skills

▶ Watch (3:29)

API access is not enough. Agents must understand the semantics of your data. Deco created a storefront skills repo from three years of optimization experience. It tells agents what to look for in logs and CDN data. One agent used these skills to find a bot running a filter explosion pattern. The bot was hidden across two systems. The agent surfaced the problem in one shot. The fix was a robots.txt update and CDN blocking.

Step Two: Triggered Agents and the System Health Monitor

▶ Watch (6:48)

Deco built a system health agent that checks CDN data and HyperDX error logs every 2 minutes. It detects latency spikes for each customer. When it finds a problem, it posts alerts to Discord and Linear. Support engineer Alini now reports 90% of her work uses the skills repo. She fixes bugs 2.5x faster per week because the agent does the investigative work.

Step Three: Agent Collaboration Toward Autonomy

▶ Watch (7:34)

The system health agent posts a GitHub issue. A developer agent reads the issue and proposes a pull request. Humans still merge the PRs. Rodrigues calls this the holy grail: specialized agents collaborating to fix problems without human direction. The key is creating small agents that do one part of the job well and then helping them collaborate.

Q&A

When do you use an agent instead of a traditional deterministic workflow? Agents are better when the reasoning path cannot be coded into a heuristic. ▶ Watch (15:52)

Can you explain the human-in-the-loop for the product reordering agent? The marketing team sets criteria with an MCP app, previews the results, then puts it on autopilot. The agent runs an ML algorithm every hour to reorder collections. ▶ Watch (20:28)

Notable Quotes

“It’s not only about giving access to an API, but your API with your data will have some formats, will have some semantics and you need to teach that to agents too.” Guilherme Rodrigues · ▶ Watch (4:11)

“People get frustrated very fast when the agents don’t one-shot any solution. And what we’re seeing in our real life is that it’s much more about creating small agents to do one part of the job well and then helping them collaborate.” Guilherme Rodrigues · ▶ Watch (8:23)

Key Takeaways

  • Centralize tools in MCP servers before giving agents any autonomy.
  • Teach agents your domain skills, not just API calls.
  • Start with triggered monitoring agents, then let them collaborate.

About the Speaker(s)

Guilherme Rodrigues is co-founder & CEO of decocms.com, an open-source framework for building and deploying MCP-based Internal AI Platforms. He spent 9 years at VTEX through its NYSE IPO, where he led the first version of the Store Framework and VTEX IO Developer Platform. He is based in Rio de Janeiro.