Community Catchup #5

The fifth weekly community catchup ran long in the best way: fast-moving model news, a serious thread on agent operations, and several demos that showed how quickly people are turning rough workflows into useful internal tools.

Highlights

The group opened with the kind of conversation that already feels normal in this community: Opus 4.7 and GPT 5.5 both landed within days, and everyone was already comparing where each fits. The bigger point was not which model wins, but how quickly the ground keeps shifting underneath every workflow.

Costs, Compute, and the Return of Discipline

Inference costs and compute scarcity were front of mind. The group discussed reports of 5-10x cost pressure, Anthropic’s compute crunch, big cloud deals, energy demand, and the possibility that subscription limits will keep tightening as usage grows.

That led naturally into routing: use the best model when the work justifies it, but do not make frontier models do every step. Smaller models, local hardware, OpenRouter-style councils, deterministic scripts, and better evals all become part of the toolkit when tokens stop feeling free.

Peeyoosh’s Tidy Bot

Peeyoosh gave a thoughtful walkthrough of what started as a challenge to build on Replit from a mobile phone for an hour a day. The first version proxied a website and let clients annotate issues directly on the page. Then it grew: WCAG audits, accessibility task lists, visual traffic maps, readability analysis, Google Analytics views, session timelines, and golden-path leakage reports.

The most important lesson was not the feature list — it was the quality bar. After auditing the Replit-generated code locally with Claude Code, Peeyoosh found exactly what many builders now see: the agent delivered fast, but took shortcuts on security and correctness. His rule of thumb: build your test harness early, test often, and think like a hacker.

Claire Puts AI Inside the CRM

Claire showed how she used Claude Code to inspect a Pipedrive instance with almost no documentation and turn it into a useful conversation with a sales director. The agent surfaced how reps were using the system, whether features like email integration were actually adopted, and where the CRM was hiding the real operational pain.

Within an hour, she had a dashboard prototype good enough to test priorities with the business. The winning insight was not a giant executive dashboard — it was a focused follow-up view that would help reps see who they needed to call next, while exposing the deeper source-of-truth problem between Pipedrive and the ERP.

Agent Sprawl Becomes Real

Ryan introduced the phrase that stuck: AI sprawl. If every team can spin up agents, databases, scheduled tasks, and copies of core data, companies will recreate the old email-silo problem with much higher stakes. The group discussed Agent Operations Center patterns, shared agent catalogs, support tiers, ownership, and cloud-hosted critical processes instead of workflows quietly depending on someone’s laptop.

Patrick’s “Annoying CEO” agent fit neatly into that theme: a backpressure agent that checks whether other agents actually did what they claimed, pokes holes in PRs, and refuses to accept hand-wavy completion.

Local Events Start Spreading

Adam closed with plans to broaden the community beyond the weekly online catchup. Online Claude events are being explored, the first town-hall style pop-up is planned for Waiheke, Wellington planning is underway, and Simon has AUT lined up as a larger Auckland venue for early July with government and AI transformation speakers in the mix.