Personal email triage agent

My personal inbox is the last queue I still run by hand. Sender-based triage got me part way. SaneBox flags things @Action or @Later, but flagged mail just sinks and never comes back up, so it piled into a couple hundred stuck threads. I do not need more folders. I need something that reads a message and tells me what it actually wants me to do.

How it works

The design is a Claude loop over Gmail that classifies each thread on a small taxonomy: disposition, action-type (reply, sign, pay, decide, schedule), and the axis that does the real work, handle-ability (auto, draft, assist, or you-only). That last one is what turns sorting into action-surfacing. The agent does not just file the thread, it tells me what it can take off my plate. It reuses plumbing I already have. The Gmail connector stays read, label, and draft only (a hard safety rail, no autonomous sends), launchd on the mini runs it, and handle-able items surface to an approve dashboard behind the same access gate.

Two rules are load-bearing. Anything touching money or a signature is always you-only. And a training-rules file lets it learn from my corrections each run, the same behavioral-memory pattern the business agents use, so it sharpens the more I correct it.

Where it stands

Planned, taxonomy-first. I am nailing down the classification before I run any live triage, and the rollout is deliberately phased: read-only digest, then supervised labeling, then draft replies, then cross-platform actions. Draft-only until it earns trust.

This is the thesis aimed at myself. The exception rate falls when an agent absorbs the routine and routes only the genuine judgment calls up to me. My inbox becomes one more department with a single head, me, and an agent doing the sorting underneath.