An agent’s Overview

Open an agent — from the Agents list or the switcher — and this is the page it lands on, the first of its seven pages: Overview, Activity, Profile, Knowledge, Skills, Tools and Settings. (A Team Agent has six: it keeps no Activity.) Everything on them is about that one agent: no number covers your account, which is why none carries an “across all 3 of your agents” line the way your Dashboard does.

The Overview is the agent drawn as one calm diagram: what it learns from on the left, the agent in the middle, and what it can do and where it answers on the right. Press a card and that part’s sheet opens beside the diagram, with every control it had. Waiting questions sit above the diagram, and the setup line shows only while setup is unfinished.

Each quick action opens the same form its own page uses, so the diagram never becomes a second editor. While a train runs, the Knowledge card’s three marks show Reading, Learning and Going live, and the lines glow on the side that is working; when it lands, a tick shows once on the Knowledge and agent cards. A train that stopped says why on the agent’s card, with Try again. While changes wait, the Knowledge card’s foot reads “2 waiting” with Train now.

Three of the panels below — the import, where this agent stands and Your hiy — carry a How this works link that opens the matching section of this page in a new tab, so reading about a panel never costs you the thing you were half way through.

The header#

Three things sit at the top of every agent page, so you never hunt for them:

• Update agent. It reads “3 changes to apply · Update agent” while something waits, “Training · Reading” while a train runs, “Up to date” when nothing waits and “Train failed” when the last one did not finish. Pressing it opens the Train sheet, with the run and what the last train learned.
• The state menu. One word says whether the agent is answering people: Live, Off, In review or Declined. Its menu holds Live, Approve every reply (a locked Preview for a Support Agent: you would approve each reply before it sends) and Off, which asks you to confirm in the menu. Below those sit the address rows for its kind: Copy link and Open hiy.ai/… for a personal twin, Get the embed for a Support Agent, Invite a colleague for a Team Agent.
• Test. It opens the dock, Test · Coach, beside the page, on every agent page. See Testing and Coach. A card on the Overview glows while a Coach change to it is open.

Under the agent’s name in the side list, v1 · History says which version it answers from. It opens when it was trained and the two histories: Knowledge changes, and Settings history, a locked Preview.

While your import runs#

Pick your pages on the create flow’s third screen and the import starts there — you don’t wait on it, and the flow doesn’t hold you. It reports on itself here, and this panel is the only place that shows both halves of it.

There are two jobs, the first one finishing is not “ready”, and the second one waits for you. An import used to start indexing by itself the moment the last page landed. It doesn’t any more: a site holds plenty you would not want your agent answering from, and the moment to say so is when you can see what actually came in. So reading is automatic and indexing is yours to start. The one exception is the import that made the agent: its first train starts by itself when the read ends (see A new agent trains itself), so that import goes from reading straight to indexing, and normally doesn’t stop at Ready when you are.

Reading fast does not mean ready soon. On a real import, the reading finished in about seven minutes and the indexing took another twenty-four. The panel never says “indexed” until the second job has actually landed — that is the whole reason it shows two phases instead of one.

You can close the page. Both jobs run on our servers. Come back and this panel picks up where it is; it is not a progress bar that only exists while you watch it.

“Ready when you are” is normal, not an error. It is where every import ends: the pages are stored and counted, and they go in when you say so. (It is also what you see when this month’s trains are used up — same panel, same pages waiting, and Update agent tells you which of the two you are looking at.)

Walking away doesn’t lose it. An import you never came back to is stored and counted like any other waiting change, so nothing is thrown away because a tab was closed. It stays out of the index until a train runs — the one you press, or the overnight one if you have switched Train overnight on.

What didn’t make it#

Not every page comes back. A page whose text is drawn by JavaScript often has nothing readable in the HTML we fetch; a dead link answers nothing at all. Any page that didn’t make it is listed here with its address and a short reason — grouped, so eleven empty pages read as one fact about your site rather than eleven separate mysteries.

Try them again re-runs just those pages as a fresh import, which waits to be trained like any other import does. Once a page has been added — by the retry, or by hand — it stops being listed. The panel stays for a day after the import finishes so a failure is still there the next morning, then Knowledge and Train own it.

What you said you’d also bring#

Setting the agent up asks where else your material lives. A site or a channel is an address we can fetch, so those two are the import above. Files, what your assistant knows about you, and your lists can’t be fetched for you — a file needs uploading, an assistant export needs you to review it fact by fact before anything is kept, and a list is rows you type.

So they’re asked for here instead, right under the import, with a button straight to the place each one happens. Each disappears the moment you’ve brought it, and if you didn’t mention any of them there’s nothing here at all.

The setup bar#

While setup is unfinished, one line sits above the diagram: Set up · 3 of 6, a small bar and the next step as its one press. It covers the six things that turn a new agent into one people are actually talking to.

It’s a checklist, not a score. Every point is attached to a step you can name. hiy doesn’t grade how good your agent is out of 100 — there’s no honest denominator for that, which is why where this agent stands is a sentence and not a number. If it can’t be read, no line is drawn: a figure nobody could load is not a zero.

A source counts when it has been read, not when it’s queued. An import creates its rows the moment it starts, long before a single page has come back, so the source step waits for the read to land and says Reading your pages while it does. The same rule as publishing, for the same reason: the step is there to tell you what your agent can actually do, and it can do nothing with a page nobody has fetched yet. (Reading is only the first of the two jobs — the step ticks after the read, and the import panel above still has indexing to do.)

When all six are done the line is gone and never comes back; the receipt, Set up · 6 of 6, moves into the agent card’s sheet.

Set it loose#

The same six stages as a list live in the agent card’s sheet, with the first unfinished one open and a button that goes straight to where it happens — the setup line’s one press is that same step. While nothing is trained yet the list says so instead of offering the next step: press Update agent in the header and the steps open up. The exception is an agent whose sources are all held back: there, training cannot help until one is ticked back on, so the list says that instead.

While your first import is still being read, Add one strong source stays open and says Reading your pages — this ticks when the read lands — with nothing to press beside it, because the read is what ticks it and there is no button that does. Nothing is wrong: the rows are there, they just aren’t answerable yet, and a step that ticked while your agent still knew nothing would be telling you to move on.

Add knowledge — on the Knowledge card’s sheet and in the teach queue below — opens this agent’s own add form, not the account-wide Knowledge library. The library has an Add of its own, but what it adds belongs to no agent until one picks it up — so material added there wouldn’t answer the step that sent you, and the agent you’re standing on would still be waiting for a source.

If the stages couldn’t be loaded, the panel says so in one line rather than vanishing — a missing next step you can’t see is worse than an error you can.

Where this agent stands#

The agent card’s sheet opens on a plain reading — nothing learned yet, not published, learning your voice, getting there or in good shape — with its ? and the next step as a button.

Your hiy#

Live, Off and the address are in the state menu; where it shows up is under Settings › Where it answers.

The teach queue#

Waiting questions sit above the diagram under Needs you: up to two one-line plates, then Show 4 more, and All questions. With nothing waiting, the block is not there. Unanswered visitor questions land here. Answer one in your own words and it becomes an authoritative correction your agent uses from then on; dismiss it and the question closes without one. Every agent’s questions, including the ones you dismissed or answered, are on the Unanswered page, grouped by what was asked, where a dismissed one can be reopened. Undo on a dismissed question puts it straight back, and All questions under the queue opens Unanswered for this agent.

It lives on this page, not on your Dashboard, and that’s a correctness fix rather than a tidy-up. Answering writes a correction against this agent. On an account-wide page a pooled queue would have taught whichever agent happened to be selected — so the queue sits on the page that names the agent it belongs to. Your Dashboard’s Needs you queue still lists what’s waiting on every agent; each row opens that agent’s page, here.

On Home in the new app every question names its agent, and Teach it writes to that agent, so one list can hold several agents’ questions safely.