What "daily triage" means here

A daily triage routine is a small, scheduled job that runs at the start of the day. Its only job is to look at everything that has landed since the last triage and answer two questions:

1. What's actually new and worth a human's attention? 2. What can be deferred, dismissed, or batched?

That's the whole thing. A triage routine is not a research workflow, not a planning workflow, and not an "AI summaries the news" workflow. It's a gate. Things either pass through the gate or they don't.

Why bother

The reason most long-running agents feel chaotic is that everything lands at once. Inbox, calendar, scheduled jobs, sub-agent results, operator messages — all of it stacked in the same stream. Without a triage step, the agent has to make judgment calls about importance at the moment something arrives. That's a bad place to make those calls.

A daily triage routine moves the importance judgment out of the real-time stream and into a dedicated, lower-pressure moment. The agent gets one short, focused task per day: sort what arrived, flag what matters, defer the rest. That's a much smaller cognitive load.

What goes in a triage routine

The minimum useful triage routine has four steps:

1. List the inputs. Pull everything that has changed since the last triage into one view. New emails, new calendar events, new scheduled job runs, new sub-agent results, new messages from other systems. 2. Tag by category. For each item, label it: urgent, defer, dismiss, batch. Categories can be refined, but four is enough to start. 3. Promote urgent to a short summary. Write one or two sentences per urgent item. The summary is what the human sees first. 4. Log everything. Save the triage output to memory or to a dated file. Next triage, the agent reads the previous one first.

That's it. The first three steps take most of the time. The fourth is the one that turns the routine into a habit.

What the output looks like

A good triage output has these properties:

  • Short. Under 500 words for the summary. A triage that goes

over 500 words has stopped being triage and started being analysis.

  • Categorized. Each item tagged urgent / defer / dismiss /

batch. The tags make the next decision obvious.

  • Actionable. Each urgent item has a next step attached. "Mike

needs to see this" is not an action; "send Mike a one-line summary of this and ask for a yes/no" is.

  • Stored. Saved to a place the next triage can read.

A short summary, written in the human's voice, that the human can read in 60 seconds.

What the routine is not

A triage routine is not:

  • A morning briefing. A briefing assumes the human reads it. A

triage assumes most of it gets dismissed. Different shapes.

  • A research report. Research goes deep on one thing. Triage

goes shallow on many things. Different intents.

  • A planning tool. Planning is about what to do next. Triage is

about what already happened. Different times.

  • An agent swarm. Triage is one focused task. Swarms are for

fan-out work.

The boundary is sharp: triage sorts. It does not act, plan, or summarize in depth.

When the routine fails

A triage routine fails when it becomes a fake gate. Common failure modes:

  • Everything is urgent. The tags have collapsed. Reset the

threshold — almost nothing is urgent, even if it feels that way.

  • Nothing is urgent. The triage has drifted into a daily

mirror of "what's on the calendar." Add a real source — the inbox, sub-agent results — and try again.

  • The human never reads the triage. The triage is now noise.

Either the human agrees to read it (and the routine stays) or the routine is removed (and the next-best thing takes its place).

  • The triage takes more than 10 minutes. The scope has crept.

Cut the inputs back to the four most useful.

How to ship one in a day

The smallest viable triage routine:

`` 1. Cron at 08:00 every weekday. 2. Pull: inbox, calendar for today, last 24h of sub-agent results. 3. For each item: assign urgent | defer | dismiss | batch. 4. Write summary to memory/YYYY-MM-DD-triage.md. 5. Send a one-message chat to the human with the urgent items. ``

That is the entire pattern. From here, the routine grows in two directions only:

  • Better inputs. Add more sources as they prove useful. Don't

add them speculatively.

  • Better summaries. Make the urgent-summary shape match what

the human actually reads. Don't make it longer.

If a triage routine doesn't ship in a day, it won't ship at all.

A small ritual that helps

At the end of each triage, write one sentence in the memory log:

> "Triage done. N items, K urgent, the rest deferred or dismissed."

That single line of meta-data turns into the most useful triage audit log you can have. Six months from now, you can ask the agent "what was urgent last Tuesday?" and it will know.