Skip to content
Skip to main content

Reach inbox zero with an AI agent

The hard part of inbox zero isn’t the cleaning — it’s the daily cleaning. Most people stick with it for a week and then quit when the queue fills back up. An interactive agent flips the calculus: the agent does the sorting and drafting, you spend five minutes approving. The 50-message backlog from yesterday is empty before your second coffee.

This recipe is the interactive cousin of the triage agent — same four-bucket model, but with a human in the loop.

The four-step flow

Section titled “The four-step flow”

1. Pull a manageable batch.

nylas email list --unread --limit 50 --json

50 is the sweet spot. Smaller and you’ll feel like the agent isn’t doing much. Larger and you’ll exceed the LLM context budget and the approval review will get tedious.

2. Categorize into four buckets.

BucketReply window
UrgentWithin hours — client issue, manager request
Action requiredToday — meeting follow-up, review
FYINo response — newsletter, status, shared doc
ArchiveNow — marketing, automated alerts

The agent shows you a summary table:

Urgent (3)         Action (8)       FYI (24)       Archive (15)
────────────       ──────────       ────────       ─────────────
Ada / Q2 plan      Rin / Review     Eng team / …   Newsletter / …
…                  …                …              …

Audit the categories before going further. If “FYI” should have been “Action”, recategorize at this step — the agent will draft what you correct it to.

3. Draft replies for Urgent and Action.

The agent generates a draft per item. You see each one:

TO:      [email protected]
RE:      Q2 plan
DRAFT:
  Hey Ada — let me block 30 minutes tomorrow morning to walk through this.
  I can offer 9am or 11am PT — what works?

You can edit, approve, or skip. Skipped drafts stay in the queue and you can revisit them at the end.

4. Execute approved actions.

After your approvals:

  • Approved drafts get sent (nylas email send --yes).
  • Items marked Archive get archived (nylas email archive <id>).
  • FYI items are left untouched.

What the agent never does

Section titled “What the agent never does”

Never send without explicit approval. This is the rule. The agent drafts; the human ships. Even if a draft is obviously fine, the click matters — it’s the difference between “AI wrote this” and “I wrote this with help”.

Customize the rules over time

Section titled “Customize the rules over time”

The first run will misclassify some messages. Encode the corrections as ongoing rules:

# Inbox-zero rules


always_fyi:
  - "from: sales@*"
  - "from: noreply@*"
  - "subject: ^\\[GitHub\\]"


always_urgent:
  - "from: *@board.example.com"
  - "subject: \\b(p0|incident|outage)\\b"

Drop these in your skill or your agent prompt. Each pass gets faster as the agent learns your context.

Daily habit

Section titled “Daily habit”

The compounding interest of inbox zero comes from doing it every day. Five minutes a day is sustainable. A two-hour purge once a month is not.

A pattern that works:

  • 8:30 AM — agent run inbox-zero while the coffee brews
  • 8:35 AM — done, all action items in your drafts folder, noise archived
  • The rest of the day — the inbox only has new mail, and you can decide in real-time whether it’s urgent

Run it from the CLI

Section titled “Run it from the CLI”

If you’re not using Manus, the same flow works as a Python script driving the CLI directly. The shape:

unread = fetch_unread(limit=50)
buckets = classify_all(unread)              # 4-bucket categorization
print_summary_table(buckets)
for msg in input_corrections(buckets):      # interactive correction
    pass
drafts = [draft_reply(m) for m in buckets["URGENT"] + buckets["ACTION"]]
for draft in interactive_approval(drafts):  # one-by-one Y/N/edit
    if draft.approved:
        send(draft)
for msg in buckets["ARCHIVE"]:
    archive(msg)

The interactive bit is what differentiates this from the cron-driven triage agent. Both share the four-bucket model.

Things to know

Section titled “Things to know”
  • 20–50 per session is the sweet spot. Below 20 you waste setup overhead. Above 50 you exhaust both your patience and the LLM context.
  • Multiple passes for backlogs. If you’re starting from 800 unread, run the loop a few times rather than asking the agent to handle the lot at once.
  • Custom buckets later. Some teams add a fifth “delegate” bucket that auto-cc’s a teammate. Doable; do it after the four-bucket version is bedded in.

Next steps

Section titled “Next steps”