Community Catchup #7

The seventh community catchup was a smaller, highly technical session: 71 minutes of workflow notes, agent design, data engineering, and practical lessons from people spending serious time inside Claude, Codex, and their own automation stacks.

Highlights

The big theme was how the job changes when the AI writes most of the code. The group kept coming back to the same question: if agents can build quickly, how do you plan, review, split work, manage context, and keep enough control to ship the right thing?

The Workflow Is the Work

Steve opened with the observation that coding is no longer the bottleneck — the workflow is. That sparked a broad conversation about planning tools, Matt Pocock-style “grill me” prompts, vertical slices, sub-agents, and whether strict specs help or accidentally hide what the agent is really building.

Adam argued for a middle ground: enough planning to keep direction, but not so much that the human disappears from the loop. Jake added one of the most useful concrete tips of the session: explore options, make the decision, then proactively compact the conversation with a clear note about what comes next. That turns a bloated exploratory context into a focused implementation session.

Adrian Turns Warehousing Into a Button

Adrian demoed an automated data warehouse builder that starts with raw source tables, profiles every column, infers relationships using statistics, asks an LLM to design the star schema, then generates and runs the dbt needed to build the warehouse.

The demo went from seven source tables to 23 warehouse tables, including derived dimensions that would usually require a data engineer to spot. Adrian was clear that the domain expertise still matters — the AI is not replacing 20 years of data warehousing judgement — but it can compress a large amount of mechanical work from months to minutes.

Triage Before Autonomy

Adam showed a customer support and engineering triage system built around Linear and the Claude Agents SDK. The first step is deliberately not “go fix it”. Instead, each incoming issue gets analysed: which customer is it, which repositories matter, is it a bug or feature request, what data can be checked safely, and how big is the likely change?

That led into the obvious next step: a follow-on agent that creates the pull request, another that reviews it, and another that tests it. The group agreed this is where the space is heading — but also where visibility and control become critical.

Jake Heads to Anthropic

Jake shared that he is joining Anthropic’s founding Australia team, working in applied AI with large organisations. The interview process was more human than expected, though transcripts and Claude-assisted feedback were part of it.

That naturally led to a discussion about how Anthropic engineers are actually working: less like 12-hour terminal binges, more like planning, dispatching jobs, reviewing outputs, and running agents asynchronously. The group talked about the dopamine hit of watching code stream by — and the need to design workflows that let people step away.

The Async Agent Future

The closing conversation circled around the idea of many agents running in parallel. Claude desktop, phone control, background sessions, Codex review inside Claude Code, and agent client protocols all came up as pieces of the same pattern.

The takeaway: the tooling is moving from interactive pair-programming toward orchestration. The human role becomes setting objectives, structuring the work, checking outputs, and knowing when to intervene.