Agents

An agent is the setup behind an answer. It decides how widely your documents are
searched, which excerpts survive, what instruction the model receives, how much of the
conversation is replayed, and which model writes the reply.

You reach them from the Agents tab in the navigation bar, or from the Agents card
on your dashboard.

Two kinds of agent

The list holds both, platform agents first.

• 🔒 Platform agents belong to Verbatim AI and are shared by every organization. You
  can read them and use them, but not change or delete them. One carries a Default
  badge: it is the agent that answers when a question names none.
• ⭐ Your agents are yours — create, tune and delete them freely.

The agent answering you right now carries an Active chip, next to the Default badge
when the two are the same agent. A card also carries the model the agent answers with —
the provider's logo and the model's name — whenever the agent pins one rather than following
the model we maintain.

Tap a card to open the agent, which shows everything it is set to. The Agents list only
ever reads: which agent answers you is decided on the agent's own page, by a button that
says so.

Choosing the one that answers you 🎯

Creating an agent does not put it to work — you pick which one answers you, and the pick is
yours rather than any one conversation's.

Three places write that choice:

• An agent's own page, opened from the Agents list: Use this agent, pinned to the
  bottom. The agent already answering shows No more use there instead, which hands your
  questions back to the default one. On the Default agent while you have picked nothing,
  that bar holds no button at all — just “This agent answers your questions”, because there
  is nothing to hand back to.
• The 🤖 button in the header of a Story.
• The Agent section of your profile.

All three write the same choice. The last two open a picker listing every agent you can use,
the agent carrying the Default badge among them, each card showing its name, what it is
for, the model it answers with and its reasoning effort. The one answering you right now carries a tick. Tap another and
it takes effect immediately.

The Agents list used to be a picker too, so a tap on a card changed the agent behind every
later question. That made it easy to change by accident, and it now takes opening the agent
and pressing a button.

There is no separate "no agent" option, because there is no such thing: until you pick, the
agent with the Default badge is the one answering. Going back to it is simply choosing
it — or pressing No more use on the agent you are on.

How long a choice lasts

It holds for the conversation you are in and the ones after it, and it survives closing
the app — you are not re-picking every morning. The 🤖 button is lit while you are on an
agent of your own and plain while you are on the default, and its tooltip names whichever
is answering.

Answers already written keep the agent that produced them, so a conversation stays readable
as it happened. And if you delete the agent you had picked, you go back to the default one —
an agent that no longer exists cannot answer anything.

Creating one

The + button opens a new agent. Only a name is required. Everything else starts on
Verbatim AI's own settings, so an agent saved with nothing but a name behaves exactly like
the default agent — and is free to drift from it whenever you are ready. That is the
recommended way to start.

Forking a platform agent 🍴

A platform agent cannot be edited — but it can be the place you start from. Open one from
the list and use the Fork button at the bottom of the page.

The create form arrives carrying every setting of the agent you forked, with one
field left empty: the name. The copy is a new agent and needs one of its own, and two
agents answering to the same name would be impossible to tell apart in the list or the
picker.

From there it is an ordinary create. Change as much or as little as you like, then save —
your copy joins your agents, fully editable even though the original never was, and
the agent you copied is left exactly as it was.

The button appears on platform agents only. Your own agents already have Edit, which
is the thing a fork would otherwise be standing in for.

What you can tune

Identity

A name to recognise it by, and an optional description of what it is for.

Retrieval

• Chunks retrieved — how many document excerpts the search brings back. Wider means
  more to draw on, and more noise. Between 5 and 10 is a sensible start.
• Re-rank results — put a model in front of the excerpts to sort them by relevance.
• Chunks kept — how many survive that pass.
• Re-ranking model — the model doing the sorting, chosen from the same list as the
  answering model. Re-ranking needs one; without it the switch has nothing to run, and the
  form will say so.

System instruction

Three texts, joined in this order into the single instruction the model reads:

• Context — what the model is looking at.
• Behaviour — how it should act.
• Tone — how the answer should sound.

Conversation

• Replay the conversation — on, earlier messages come back into context; off, every
  question is answered on its own.
• Messages replayed — how many of the latest ones. Leave it empty to replay the whole
  conversation.

Model

• Answering model — the model that writes the reply.
• Reasoning effort — Fast or Thorough. Thorough is slower but handles harder
  questions.
• Temperature — how much the model may vary its wording. The lower it is, the more
  repeatable the answer.

Both model fields open the same picker: the models Verbatim AI serves, each with its
provider's logo, its name and a line on what it is good for, and Platform default at the
top. The description is the part worth reading — it is what tells a quick, inexpensive model
apart from one that costs more per question but handles nuance better. The agent's page
names its models the same way, so what you picked is what you read back.

Platform defaults, and what "empty" means 🌱

Every optional setting reads Platform default until you pin a value. That is not the
same as blank: it means the agent follows a value Verbatim AI maintains, and moves with
it when we improve it.

The consequence worth knowing: clearing a field puts it back to following the platform,
it does not blank it. That is usually what you want when a value you pinned earlier is no
longer serving you.

Some settings can never be cleared, because an agent always needs one — the name, the
number of chunks retrieved, the two switches, and the reasoning effort.

Editing and deleting

Open an agent by tapping its card, then use the Edit button at the bottom
of the page — the same corner a
platform agent offers Fork in, since only one of the two ever applies. Platform agents
have no Edit action, and the edit page refuses them outright.

The 🗑️ bin in the header of the edit page deletes the agent for good, after a typed
confirmation.
Deleting does not rewrite the past: answers already produced with the agent keep naming it,
so conversations stay readable exactly as they happened. And if it was the agent you had
picked, you go back to the default one from then on.