Custom agents

The easiest way to picture a custom agent is as a new colleague you are onboarding. You tell them what their job is, give them access to the data they need for it, hand over the materials and agree what they should do regularly. From then on they do it themselves.

This is not just an analogy – it is genuinely how an agent is built in Datify. It has its own data access, not one borrowed from you, exactly as a new colleague gets their own login rather than yours.

It pays off wherever the same way of working repeats: reviewing contracts, sorting enquiries, preparing documents the way your company does them.

Onboarding an agent

The sections in the agent settings match what you would go through with a new starter:

What you cover with a colleague and where it lives on an agent

When onboarding a colleague	On an agent
You tell them what their job is and how you like it done.	Instructions
You arrange access to the systems they need for it.	Knowledge
You hand over materials – a price list, a template, a methodology.	Files
You agree what happens every Monday and what at period close.	Tasks
You decide who will work with them.	Access
You give them a ceiling for spending without asking.	Spending

In one respect it differs from a colleague fundamentally: it will not fill in the gaps or ask the person at the next desk. Whatever you do not give it in the instructions and the knowledge, it has nowhere to get. So it pays to be more specific with an agent than you would be with a person.

Knowledge belongs to the agent

This is the one thing worth understanding before anything else. Stick with the colleague picture: the agent gets its own access badge, not a copy of yours.

It therefore works with exactly the datasets you assign to it, not with what you can see. Two consequences follow, and they complement each other:

• If you assign no dataset, the agent has no access to any workspace data. It does not inherit yours. It can still chat, use other tools and run from tasks.
• If you do assign one, someone who cannot see that data themselves can get an answer from it through the agent. That is by design, not a gap: the agent is a separate identity that you lend to people.

A user's direct data access does not change because of this. If you click a link in an answer pointing to a table you have no permission for, it will not open. Agent knowledge and access covers this in detail, including roles.

How it differs from the default agent

If a custom agent is a colleague, the default agent is your own note about how the assistant should talk to you. It is the one you talk to when you pick none. It applies only to your conversations and nobody else sees it. You can change its name, description, instructions and model, so write into it what it should always take into account.

But it has none of what makes a custom agent a colleague: no datasets, no sharing, no tasks and no versions. It has no badge of its own either – it works with your permissions. That is also why it does not appear in the agent overview.

Where to find your agents

Every agent you have access to lives on the Agents page. Open it from the sidebar with All agents, and once your agents stop fitting into the panel, with the More tile as well.

The page offers two layouts – List and Cards – and remembers your choice. Ordering is under Sort:

What you can sort by

Sort	What it is for
Custom order	The default. Drag agents wherever you want – the order is yours and does not change for colleagues.
Name	Alphabetical A–Z or Z–A.
Last executed	Who ran most recently and who has been quiet.
Runs (30d)	Who actually gets used.
AI credits (30d)	Who costs what.

Dragging works only in Custom order; as soon as you sort by a column, the handle disappears.

Each agent shows its description (or No description), its managers and its activity over the last thirty days. An agent you have no access to is marked You don't have direct access to this agent. – such a row will not open.

How to create an agent

Create one with New agent. There are two routes:

Two routes to a new agent

GuidedDesign a new agent

You describe in your own words what the agent should do. Datify turns it into a draft that you adjust in the Review your agent step.

ManualCreate agent manually

You write the name, description and instructions yourself, without a draft.

Whichever you pick, the agent is created only when you confirm it – while you are still adjusting the draft nothing is saved, and rewriting your brief repeatedly produces no duplicates. Whoever creates an agent becomes its first manager.

The name needs at least two characters and fits into fifty, the description into two hundred. You cannot save an agent without instructions; Datify writes Instructions are required. That is deliberate: an agent without instructions is just the default agent under another name.

What you configure on an agent

The settings have two parts and the difference is worth remembering.

The Agent settings block holds Instructions, Files, Knowledge and Tasks together. These four form one version – every save creates a new one and finished runs stay on theirs. Agent activity and versions describes it.

Outside the block sit Access (who may use and manage the agent) and Spending (a credit limit for automated runs).

Until you save, the panel keeps a counter of unsaved changes, with Discard and Save beside it.

How to write instructions

Instructions are the agent's job description, not today's task. A task is different every time you type it into the chat; instructions say how the agent should always work. The same rule applies as for briefs: specific beats general.

Write them the way you would onboard someone who has never worked with you and will not ask any follow-up questions. Good instructions say four things:

1. What the agent is. "You are an assistant to the sales department."
2. Where it takes data from. "Take enquiries from the Enquiries table, prices from the Price list."
3. What the output should look like. "Answer briefly, in bullets, amounts always in CZK."
4. When it should ask. "When you are unsure about the category, ask instead of guessing."

Bad instructions describe a one-off task ("summarise my enquiries"). That is like hiring a colleague for a single activity – you would need another one for everything else. Such a brief belongs in the chat, or in a task.

Model and files

For the model, Latest means the agent always runs on the newest available model without you having to watch it. Pick a specific model when the agent should do something consistently the same way. Your choice in the chat will not override the agent's – Choosing a model describes it.

In the Files section you attach material the agent should have at hand on every run – a price list, a template, a methodology. Ten of them fit, and the same allowance is shared with files attached to individual tasks. The source stays unchanged: when the agent works with it, it creates the result as a new artifact.

Duplicating and deleting

The menu holds Duplicate and Delete. The copy is created with a number in its name and is independent of the original.

Deleting is confirmed by a dialog and removes all of the agent's permissions too.

Related articles