Skip to content

Agentic weather

How the agent decides when to act, and when to ask you

Jason Deese, Founder & Chief Meteorologist

Published · 7 min read

Flash on a laptop and a phone, the surfaces where Flash Agent asks for permission before it acts

When it acts, and when it asks

In this article · 6 sections

An agent that can move a practice or reschedule a crane lift is only useful if you know exactly where its authority ends. Flash Agent draws that line in one place: it answers, notifies and prepares changes on its own, and a person confirms before any schedule or record changes.

The line: what it does alone, and what waits for you

  • On its own

    Answers
    Answers questions in plain language, citing the cell and model run
    Notifications
    Posts the call and the reason in Slack or Teams; notifies the people concerned
    Schedules
    Prepares the change: the moved practice, the rescheduled lift, the shotgun delay
    Records
    Prepares tags, tasks, cases, work-order holds and daily-log entries
  • Waits for a person

    Answers
    —
    Notifications
    —
    Schedules
    Every schedule change is confirmed before it is made
    Records
    Every record change is confirmed before it is made

The confirmation happens where people already are: in Slack, in Teams on an adaptive card, by SMS or in the Flash app. The person who confirms is logged with the action.

Four guardrails every action passes

  1. Permissions per connector. Each tool is switched on per site with its own scope. Turn one off and Flash Agent loses it immediately; the log keeps the history.
  2. A cited cell and model run. The 1×1 km cell id and the run time sit under every answer, so anyone can check it against the map and the accuracy method.
  3. Every action logged: who asked, what was read, what changed and who confirmed it.
  4. Human confirmation for schedule and record changes. Flash Agent prepares the change and waits.

On the Flash Agent page an illustrative proposed action passes all four in turn: move Thursday practice to 5:30 pm in Google Calendar. The calendar write is allowed for that site, the answer cites its cell and the 12:02 run, the entry is prepared, and it shows "waiting for a person" until a coach confirms.

What a permission actually covers

Permission is set per connector and per site, and each connector has a narrow job. Four kinds of access come up: reading what the connector exposes, proposing changes, posting messages, and confirming. A few examples from the connector list:

Google Calendar
Reads events on the site’s calendars. Proposes moving or shortening events inside a risk window, with a note to attendees. Every event change waits for a person.
Procore
Reads the schedule and the daily log. Proposes moving lifts and pours, logging the delay and texting the foreman. Every schedule change waits for a person.
Salesforce
Reads accounts and sites in a hail path. Proposes hail tags, tasks and cases with the cell and run attached. Every record change waits for a person.
Slack and Microsoft Teams
Read questions in the channels Flash Agent has been added to, post the call and the reason, and collect confirmations.
Webhooks and the API
Send each alert, all-clear and model run to your own endpoints. Your own rules decide what happens next; Flash Agent changes nothing in your tools through them.

Why the line sits at schedules and records

Answers and notifications are cheap to get wrong and easy to ignore. A moved practice, a rescheduled pour or an edited tee sheet changes what people do that day, and someone has to own it. Putting the confirmation exactly there keeps the agent fast on the questions that matter in the moment and keeps a named person accountable for anything that changes a plan.

It also keeps the record straight. Because every action carries the cell, the model run and the person who confirmed it, the log shows why a decision was made and by whom, alongside the alert log that already records who was warned and when play resumed.

An empty floodlit athletic field at dusk under heavy cloud, where a moved practice needs a coach to confirm it

A person confirms

The agent prepares the change. Someone with a name makes it.

What Flash Agent is not

  • It does not replace the Weather Command Center. The Command Center stays the map and the log for people who want to see every cell; both read the same cells and write to the same event log.
  • It does not forecast separately. Flash Agent, the Command Center, the mobile app and the API all read the same 1×1 km model run.
  • It is not the alert system. Lightning alerts still reach the field by app, SMS, email and horn and strobe.

Frequently asked

Not on a schedule or a record. It can answer, notify and prepare a change on its own; moving a practice, rescheduling a lift or editing a tee sheet waits for a person to confirm, and that confirmation is logged with the action.

Yes. Each connector is switched on per site with its own scope. Turning a connector off removes the agent’s access to it immediately, and the log keeps the history.

Who asked, what was read, what changed and who confirmed it, for every action.

Jason Deese

Founder & Chief Meteorologist, Flash Weather

Live demo · No sensors required for the forecast

See your own sites on the map before the next storm.

We'll load your locations, replay last season's strikes against Flash's predictions, and show you the alert log your safety officer would have received.