Calendar as a constraint graph.

Your calendar is a constraint graph CORE

Most people treat a calendar as a list of meetings. For agentic work, a list isn’t enough. A calendar is a constraint graph: a set of blocked times, windows of availability, priorities that override each other, and soft constraints (you do not take meetings before 9 a.m., your family eats dinner at 6, you keep Thursday afternoons for deep work). The agent works with the calendar well only if you give it the graph, not just the list.

Three things follow from the graph view.

Availability is derived, not listed. An agent asked “when am I free Friday?” should not just read the Friday events. It should compare Friday events against your working hours, your recurring personal blocks (the ones tagged as “soft” in your calendar), and any constraints you named in the prompt (“I do not take meetings in the hour before a class”). Free time is the absence of conflicts you count as conflicts, not just the absence of events.

Scheduling is a negotiation with yourself first. Before the agent proposes times to another person, it should propose times to you. “You have three windows that look free Friday afternoon; which do you want me to offer?” That second you is the chance to catch the window that is technically free and actually terrible — the 4:30 slot right before you pick up a sibling, the Tuesday morning block that collides with an unlisted travel day. The agent cannot see what is not in the graph.

Calendar events have side effects in the thread. Accepting a calendar invite is not a private action. It signals commitment to the person who sent it. Declining is also communication. An agent that RSVPs on your behalf — even to a “reasonable-looking” invite — has sent a message to a human who will treat the RSVP as you. This is why calendar actions fall under drafts, not sends exactly the way email replies do.

The practical version of this: you tell the agent what your calendar looks like (constraints, working hours, soft blocks), the agent reads the events, and you confirm every change.

Drafts, not sends, for calendar CORE

The drafts-not-sends rule applied to calendar work has three specific shapes.

Proposed events land on the Agent Access calendar first. When the agent creates an event based on a conversation — “let’s meet Tuesday at 3” pulled out of an email thread, a dentist appointment from a confirmation — it lands on your Agent Access calendar as a tentative event. You review, you confirm, and then you move it to your real calendar (or mark it confirmed on the Agent Access calendar if you have configured that calendar to count). The agent does not place events on your real calendar directly.

Calendar invites to other people go to drafts. When the agent prepares an invite that would send to a third party (your teacher, a friend, a mentor), it drafts the invite and lets you be the one to click Send. This is the same pattern as email replies. The invite is a message. The message does not go out without a human hand on the button.

RSVPs are yours. The agent can read a calendar invite and recommend accept, decline, or propose-new-time. The agent can even draft the propose-new-time message. The agent does not click accept or decline on your behalf. This is the one students most often relax — “but it’s just an RSVP, it’s fine” — and it is the one with the most direct relational signal. A friend whose invite you ignored for three hours because you wanted the agent to “just accept it” is a friend who felt ignored. You keep the click.

A note on recurring events. If the agent proposes a series — weekly tutoring, a standing check-in with a parent — you review and confirm the first instance and the recurrence rule. The recurrence is a commitment too; do not let the agent extend a commitment on your behalf by scheduling thirteen weeks out without your review.

The meeting-prep brief: what a useful one contains CORE

You have an upcoming meeting. You have a short window to prepare. The agent is going to help you produce a meeting-prep brief — a two-page document that pulls the scattered context around the meeting into one place so you walk in ready.

A good brief has five sections.

1. Meeting context. One paragraph. Who, when, where, how long. What is on the agenda — or what the agent inferred the agenda is from the thread. Who called the meeting and what outcome they are likely looking for. If there is a stated agenda, quote it; if there is none, note the absence honestly.

2. Recent thread history. A short summary of the last three to seven messages in the thread that led to this meeting. Names and dates. The goal is not a transcript; it is “here is the state of the conversation that the meeting is the next move in.” If the thread lives in the agent-access label, the agent can read it directly; otherwise, you may need to paste the relevant thread into the prompt.

3. Open questions. The questions still unresolved going into the meeting — from the other side (“you asked whether I can submit Friday or Monday; I have not answered yet”) and from your side (“I still do not know if the extension applies to the rubric”). Three to six questions, ranked roughly by importance.

4. Documents and references. A list of the documents the meeting is likely to reference — linked or named. A feedback document your teacher sent. A rubric. A draft you shared. If the agent can link to them, it does; if it cannot, it names them clearly so you can find them yourself.

5. What you want to leave with. One sentence. “At the end of this meeting, I want to have agreed on a submission date and a definition of ‘complete.’” This is the sentence you write yourself. The agent can draft it — it will usually draft something plausible — but you overwrite it with what you actually want, because the agent is guessing and you are not.

Two things the brief is not. It is not a script of what to say — writing one of those usually leads to a stilted meeting and the student stops using the tool. It is not a transcript of the thread — that would be longer than the meeting itself. The brief is pruned context, not exhaustive context. If you find yourself skimming your own brief in the meeting because there is too much in it, next time ask for half the length.

The meeting-prep prompt pattern CORE

Like the triage and drafting prompts, the prep prompt has a stable pattern.

1. The meeting. “I have a meeting with my teacher on Friday at 3 p.m. about my late assignment.” Name it; the agent cannot know which of your meetings you are prepping for.

2. The source material. Point the agent at the material it is allowed to read. “Read the thread in my agent-access label titled ‘Late assignment extension.’ Also pull in the rubric document linked in the third message.” If the material is outside the access surface, you paste it in directly.

3. The brief shape. “Produce a meeting-prep brief in five sections: Meeting context, Recent thread history, Open questions, Documents and references, What I want to leave with — filled in for the first four, and leave the last one blank for me to fill in.”

4. The honesty rails. “Quote the stated agenda if there is one; if there is not, say so. Do not invent quotes from the teacher or myself. If a document is referenced in the thread but you cannot access it, list it under Documents and say you could not read it. Do not guess its contents.”

5. The constraint. “Read-only for email. Do not create calendar events, do not send messages, do not modify the thread’s read state.”

Same five-part shape as the triage prompt. The one difference: the brief’s fifth section (What I want to leave with) is the one you fill in after the agent produces the rest. You own the intention; the agent prepares the context.

A secondary prompt, worth using: after the agent produces the brief, ask it “what context did you omit that a careful reader would want?” and read the answer. The agent’s own sense of its gaps is sometimes useful, and it is a good prompt-discipline habit — asking the agent to critique its own output is a cheap check that costs nothing and catches occasional real omissions.

Calendar sweeps: the weekly rhythm CORE

Meeting prep is per-meeting. A calendar sweep is the weekly rhythm that keeps the whole system honest.

A calendar sweep is a read-only agent pass over the upcoming week. The agent reports:

• Conflicts or near-conflicts — overlapping events, back-to-back meetings with no buffer, travel times that do not fit in the gaps.
• Uncommitted items — tentative events on the Agent Access calendar that have been sitting there for more than a few days. Each needs a confirm or a decline.
• Open loops — threads in the agent-access label that reference meetings not yet on the calendar. “You agreed to call your grandmother this week; no event is on the calendar for it.”
• Stale commitments — recurring events that have not happened in weeks and probably should be cleaned up.

The sweep is short — a quick read of the agent’s report, plus whatever follow-through the report surfaces. It is the equivalent of the Sunday-night calendar review a careful person does by hand, but faster. The sweep produces no outbound actions. You decide, after reading, what to act on.

Run the sweep once before the end of this lesson, as part of your capstone Entry 3 setup. Make it a weekly rhythm going forward. A calendar without a weekly sweep drifts; the sweep is how you catch the drift early.