What it involved
- An MCP server, so the agent reaches real systems through described tools instead of improvised API calls.
- Long-running workflows handed to Temporal, so work spanning minutes and retries survives a restart.
- Agent behaviour traced in LangFuse, which is what makes a wrong answer debuggable after the fact.
- Current work, so this entry says what is being built rather than what it has returned.
This one does not have a full write-up yet. If you want the detail on it, ask me directly and I will walk you through the decisions.