― HELP · THE AGENT

How to ask the agent a question

The agent fetches a forecast, checks it against your limit and shows its working. Here is where to open it, what it sees on screen, what it will not do and how to read the answer.

5 min readUpdated Checked against the product
ON THIS PAGE
  1. Steps
  2. Where to find the agent
  3. What it sees on screen
  4. What it fetches
  5. How it answers
  6. Good questions
  7. Follow-ups and starters
  8. Signed in or not
  9. Limits and problems
  10. Questions

Steps

  1. Open the panel

    On most pages press Ask the agent in the header, or the floating Ask the agent pill on content pages. In the instrument at /app there is no header button. Press Explain this in the inspector, or Ask on the card headed The agent noticed. Cab mode has no agent.

  2. Pick a starter or type

    The panel greets you, then shows up to three starter questions you can press. Or type in the box headed Ask the agent, up to 600 characters. Press Enter to send, or Shift and Enter for a new line. Send needs at least 2 characters.

    Tip: Say the place, the height and your limit, for example Clonmel, 85 m, 25 mph, 07:00 tomorrow.

  3. Watch it work

    While it works, the panel shows short working lines and a list headed What the agent checked. Each line names the step, such as counting ensemble members or checking the nearest station, with a tick or a cross and a UTC time. Only one question runs at a time.

  4. Read the answer

    The answer is short, with the source, the valid time in site time and the height for each value. Below it is an evidence line listing the sources, then who wrote it. If any number could not be traced it is marked (unverified).

    Tip: See what (unverified) means.

  5. Follow up

    Up to three follow-up chips appear under the answer, such as asking for the ensemble odds for that hour or the next hour over your limit. Press one to ask it, or type your own.

  6. Close the panel

    Press Close, press Escape or click outside the panel. A rule or record card that you did not confirm saves nothing.

01Where to find the agent

The header of the main site pages, including the fleet board and the alerts page, has an Ask the agent button. Content pages also get a floating Ask the agent pill that hides while the panel is open. The instrument at /app has its own layout with no header button, so there you press Explain this under the inspector, or Ask on the card headed The agent noticed. Both open the same panel. Cab mode has no agent. You do not need an account to ask a question.

02What it sees on screen

When you open it from the instrument, the panel is bound to what you are looking at: the place, height, limit, unit, persona, the hour on the time strip and the readings behind them. It answers about that first and fetches more only if your question needs another hour, height or place. Its starters change with your persona, for instance asking whether you can lift, fly, spray, play or go out at the hour shown. On the fleet board it reads up to 40 of your sites.

03What it fetches

Every turn, the agent fetches before it speaks and never forecasts from its own knowledge. It has tools to find a place, read the forecast at 10, 80, 120 and 180 m, read the wind at a given height, count ensemble members, check the nearest METAR or buoy, compare with ERA5 climate, walk a route, find the next hour over your limit and, when you are signed in, read your saved sites. A turn can run up to six rounds of tools, 20 seconds each. Earlier answers in the conversation are context and not data. The agent fetches again, so numbers are never reused from before.

04How it answers

Each value is given with its source, the valid time in site time and the height. Gust is always the 10 m gust and never a gust at height. Ensemble members exist only at 10, 80, 100 and 120 m, so when you ask for another height the agent says which level it used. Odds are plain member fractions, with the count of members, and are uncalibrated. If members are missing the first sentence says ENSEMBLE INCOMPLETE. Answers are plain text of about 120 words or fewer, with no headings. The agent never says it is fine to go ahead. It compares to your limit and stops.

05Good questions

Ask for a number or a comparison. For example: what is the wind at 120 m at 14:00 tomorrow in Clonmel, when is the next hour over 25 mph, how likely is it that gusts reach my limit tomorrow morning, do the observations agree with the model, or why do the models not agree. For your own sites, ask which to shut first. Five kinds of question are refused with a fixed sentence. See why the agent will not answer some questions.

06Follow-ups and starters

Starters are shown before your first question. After an answer you get up to three follow-up chips chosen from what the agent has not yet checked, such as the ensemble odds, the next hour over your limit, whether the observations agree, or the best window to lift, fly, spray or play tomorrow. They are shortcuts, and you can ignore them.

07Signed in or not

You can ask anything without signing in. Signing in lets the agent read your saved sites, compare them, and propose saving a rule, which you then confirm. It also keeps your facts and history on your account instead of in the browser. See why the agent asks you to confirm and how the agent remembers. If it cannot read your sites because you are signed out, it says you need to sign in.

08Limits and problems

The agent accepts 12 questions a minute from one connection, with no daily cap and no plan limit. If the narrating model is unavailable, an engine answers instead with a plain read of the same data. If even that fails you see Agent unavailable and the reason. See what (unverified) and ENSEMBLE INCOMPLETE mean.

Questions

Can the agent forecast from its own knowledge?

No. Every number comes from a tool it ran this turn, or from the packet on your screen. If a figure is not in either, it says it is not in the packet.

What happens when the agent cannot answer?

For the five refused kinds of question it gives one fixed sentence and does nothing else. For anything else it fetches what it can. If a tool fails you see a cross on that line in the list of what it checked, and the answer is built from what did arrive.

Does the agent give go or no-go advice?

No. It compares the forecast to the limit you set, gives the margin and the ensemble odds, and stops. The decision stays with you and your competent person on site.

Does Explain this send a question?

No. It opens the panel with your screen bound to it. Then press a starter or type your own question.

Can it change my settings?

It can save a rule on a site you already have, or freeze the packet on screen, but only after you press Confirm. It can also turn your morning brief on or off at your request, and remember a fact you tell it.