Community Catchup #4

The fourth weekly community catchup ran for a practical session on building agents that actually survive contact with real work — budgets, review loops, domain tools, and all.

Highlights

The strongest theme was using AI deliberately. As unlimited-feeling usage starts to tighten, the group compared ways to keep velocity high without burning through expensive tokens: use Opus when it matters, let Sonnet handle more routine execution, cache aggressively, convert bulky inputs to markdown, and avoid asking agents to manually shuttle data when a deterministic script would do.

Constantino’s Design-to-Web Workflow

Constantino demoed a tool that turns design inputs into live web components — exactly the kind of bridge teams have been asking for between static design systems and working prototypes. For people trying to get non-technical teams closer to real product interfaces, the appeal was obvious: fewer handoffs, faster iteration, and a path from visual intent to something usable in the browser.

The group encouraged him to consider open sourcing it if he did not want to commercialize it directly. In a world where implementation effort is shrinking fast, sharing the tool could let others contribute the missing pieces while he focuses on the parts he knows best.

Jared’s Agent Harness for Data Engineering

Jared brought a deep dive into a custom agent harness built for data engineering teams. It began before skills and MCPs were mature, so he had to build many of the pieces by hand: tool orchestration, plan tracking, sub-agent review, dynamic tool selection, session files, and visual output renderers.

The most useful pattern was the reviewer sub-agent. The main agent does the work, then a separate reviewer checks files and acceptance criteria before the step is considered complete. It is a simple idea, but it consistently catches the last few percent of issues that otherwise slip through.

The Shape of Better Agent Interfaces

A recurring point from Jared’s demo: agents should not only return text. When an agent produces SQL, an entity relationship diagram, a pipeline, or a structured plan, the interface should know how to render that object properly. Cowork’s canvas hints at where things are going, but the group agreed there is a much bigger opportunity for domain-specific visual interfaces wrapped around agents.

Token Discipline Becomes a Product Skill

The cost conversation was grounded and useful. William shared an example of Claude moving data from Clockify to Notion by literally typing values through MCP calls — including IDs and potentially bulky encoded content. The group pushed back toward a cleaner pattern: have the agent write the API bridge, then run the transfer deterministically.

That pattern kept coming up: agents are excellent at planning, translating intent, and creating glue code. But once the path is known, ordinary code is still cheaper, faster, and easier to test.

What’s Coming

Adam confirmed the weekly catchups will continue, with in-person meetups likely every couple of months. There is also interest in regular online Claude-focused sessions so the wider New Zealand community can participate without being tied to one city.