Community Catchup #13

The thirteenth community catchup focused on agents moving from solo tools into shared team environments: Slack bots, group-chat agents, adversarial review loops, automated data warehouses, cross-platform product building, and the next wave of community events.

Highlights

The through-line was AI becoming part of the operating system of a team. Not a chat window on the side, but something sitting in the same channels as the humans, with context, tools, memory, and responsibility boundaries.

Claude in the Room

Steve and Adam opened with Claude-in-Slack patterns. Adam described earlier experiments where agents had access to MongoDB, GitHub, and deployment context, then handed work between each other. In customer group chats, an agent can now understand the company context, implement a small feature, open a pull request, and let Adam review it.

That raises a new business question: if the work becomes that transparent and fast, how do you charge for the expertise that made it possible? The group leaned toward retainers and responsibility-based pricing. The human still owns the system, the guardrails, the judgement, and the risk.

Review Loops Are Becoming the Process

Adam shared how useful the phrase adversarial review has become: ask the agent to dispute the PR, find what is wrong, then fix it before merging. Dominik added a more structured version: give Opus a PRD, break the work into phases, run implementation agents, run adversarial review agents after each phase, fix issues until a confidence threshold is met, commit, and move on.

The interesting part is not just quality control. It is context control. The main agent keeps the plan and state; sub-agents do the detailed work and reviews. That pattern is quickly becoming one of the community’s default approaches for serious builds.

Safety, Open Models, and Compute Pressure

The group also touched the uncomfortable side of open models. Adam had tested small unrestricted models that would answer plainly dangerous questions, and the group discussed whether hosting stripped-safety models on open platforms helps or hurts the industry.

That thread widened into model access, Fable restrictions, chips, power, and video generation. JP and Steve noted that adult-industry money has historically accelerated streaming, 3D, and media technology — and may do the same for AI video. It was a messy but honest conversation about incentives, regulation, and where infrastructure progress often comes from.

Adrian’s Ten-Minute Data Warehouse

Adrian demoed a tool that looked like a direct attack on weeks of traditional data engineering work. It connects to sources like HubSpot, Xero, Shopify, CSVs, shapefiles, and Postgres dumps; profiles tables; infers relationships; designs warehouse models; generates DBT; runs against Snowflake or other backends; and cleans dimensions through a learning loop.

One test profiled tables in 12 seconds and built a usable warehouse in minutes. Adrian’s bigger question was the right one: if data warehouses become this cheap and fast, what use cases become possible that were never worth doing before?

Claire’s OP Spot Flywheel

Claire shared progress on OP Spot, now five weeks in and already spanning iOS, Android, and web. The web version became important once she saw the usual app-download drop-off: only around a quarter of landing-page visitors make it into the app.

Her workflow is increasingly like managing a dev team. Notion holds the product brain, Claude Code has the context, the design system keeps platforms consistent, and multiple agents run at once. When something goes wrong, she runs a retro with the AI the way she would with a human team. The goal is bigger than a map of op shops: aggregate community-shop inventory and help the whole sector level up digitally.

Impact Labs Are Getting Real

Adam closed with community updates. Chris Watson is finalising the Wellington ambassador path, and the first Claude Impact Lab is likely to kick off in Wellington around mid-August with councils, Creative HQ, Anthropic support, API credits, and real public-sector problems prepared ahead of time. Auckland and Christchurch are also on the roadmap.

The July 1 Auckland meetup at AUT is already huge: around 900 applications for a 380-person lecture theatre. The community has moved well beyond a small meetup. The challenge now is turning that energy into useful projects, repeatable practices, and real-world impact.