# The Managed Agent Routine Map

Worked example: Morning Brief Agent for a fractional COO.
One workflow. One routine. One page.

## 1. Workflow name

Morning Brief — daily, 7:00 AM local.

## 2. The manual version (before)

- Owner opens 6 tabs (email, calendar, Notion, Slack, Linear, bank).
- Pieces together "what matters today" from memory.
- Writes a 3-line note to self.
- Forgets one client follow-up about 2x/week.

## 3. Inputs the agent reads

- Calendar (today + next 48h)
- Email since last brief (filtered by sender list)
- Active client tracker (Notion DB)
- Open follow-ups queue (source-of-truth file)
- Yesterday's brief (for continuity)

## 4. Source of truth

`brief-context.md` — single markdown file. Sections:

- Active clients + current stage
- Standing priorities (this quarter)
- People to never miss (VIP list)
- Topics to flag (litigation, contract, payment)
- Decision log (last 14 days)

## 5. Agent role

Drafts a single morning brief. Never sends. Never acts.

## 6. Trigger

Cron at 06:45 local → drops draft into `inbox/morning-brief/YYYY-MM-DD.md`.

## 7. Draft output (template)

- Today's calendar with one-line stakes per meeting
- 3 priorities surfaced from source-of-truth
- Open follow-ups aged > 48h
- 2 flags ("watch this," "decide today")
- 1 question for the human

## 8. Human approval boundary

Owner reads the draft, edits, decides what to act on.
Agent never emails, messages, or schedules.
Agent never deletes from the follow-up queue — only proposes.

## 9. Escalation rule

If a flagged topic appears (litigation, contract, payment, VIP),
agent adds a 🚩 marker and a one-line "why this matters."
If confidence is low, agent says so and lists the gap.

## 10. Review rhythm

Daily: 60 seconds (read, edit, act).
Weekly: 15 minutes (update source-of-truth, prune queue, adjust prompt).
Monthly: review what got missed, tune escalation list.

## 11. What the human still owns

Every send, schedule, decision, and "this is sensitive — handle myself."

## 12. Failure modes (and the runbook)

- Missing input → agent flags, does not guess.
- Conflicting sources → agent presents both, owner chooses.
- Tone drift → owner adds an example to source-of-truth.
- Stale context → weekly review catches it.