When the obvious next step is the wrong one.
The natural extension of the Hostaway integration pointed toward a bank feed — automatically reconciling income data against actual payments, removing the last manual thread from the reporting cycle. We scoped it. And we decided not to build it.
Bank feed integrations for custom-built projects carry a level of compliance overhead that sits well outside the scope of a reporting tool. More importantly, the reconciliation UI required to communicate what an automated process has done — the edge cases, the mismatches, the exceptions — would have taken the project somewhere neither of us wanted it to go. We would have spent months building something that made the process more complicated to audit, not less.
Instead, we explored something smaller and more interesting. As we had started building AI-powered tools internally, we saw a different kind of opportunity — not a full automated reconciliation pipeline, but an agent-based approach that could handle the messy, judgment-heavy parts of reconciliation without requiring a complex UI to surface it. A Reconciliation Agent is where the thinking pointed. But we did not start by building it.
Starting with a skill, not the full agent
The first version is much smaller than the eventual shape. Rather than commit the hours a full agent build would have asked for, we started by writing a reconciliation skill — a focused, composable piece we could put in front of the actual work and iterate on cheaply. Alongside it, an HTML report template the agent could use to format and present its decision log, so the reasoning behind a reconciliation pass was legible rather than buried in a chat history.
We shared both with Aire Spaces to use through Claude Cowork. As a production setup it is not ideal. But as a testing surface it was hard to beat: easy to put in their hands at a stage when skills were still new to them.
What we realised wasn’t just that agents could change the cost and effectiveness of how we build tools for our clients. It was that we can iterate with agents inside the workflows we build for them — the same short feedback loop we aim for with everything else we build — and use that loop to find new ways through the hard problems sitting in those workflows.