On this page
The functions in this article are built and safeguarded, but have barely been used in practice — the persona suggestion and material ingest least of all. Try them first on an agent that isn't doing anything important yet, and read suggested texts carefully before you apply them.
Where all this lives and who may use it
You maintain persona, knowledge base and twin link under Agent Center → Agents: pick the agent on the left, everything else sits in its details. Only someone with the “Manage organization” permission may change anything there. With “Manage members / roles” you see the persona, tools and budget but cannot save — and the private twin knowledge stays hidden from you entirely.
There is one exception: if you yourself are linked as a digital twin, a “My twin” entry appears in your menu. There you maintain your own twin's identity, persona and private knowledge without any special permission — but the link itself is still set by managers only.
What a persona is — and what it cannot do
The persona is the core of an agent's instructions. It sits in each agent's details under “Persona (system prompt core)” and is applied with “Save as new version”. Every change creates a new version — the header shows you which version an agent is currently working from, so it stays traceable which text produced a given result.
bricksta automatically prefixes every persona with the same terse behavioural frame: that the agent is a careful colleague with a clearly bounded remit, works exclusively the tasks assigned to it, sticks strictly to its permissions, escalates to a human when in doubt and invents no process steps. You cannot switch this frame off — it applies to every agent alike, including one whose persona you wrote entirely yourself.
A persona never governs what an agent may do. Permissions are granted by you alone, through tools and budget. A text that awards itself extra authority stays without effect — no matter who wrote it or how convincing it sounds.
Writing a good persona
A persona is not role-play but a work instruction. Four things make it useful:
- The task, concretely: not “help with onboarding” but “check incoming onboarding documents for completeness and record missing items as a note on the task”.
- The tone, where it matters: for customer contact, this is where you set how formal, how detailed and in which language it writes.
- The boundaries, explicitly: what the agent should deliberately NOT do even though it could — “don't estimate unclear prices, hand the task back to sales instead”, say.
- The context it otherwise lacks: your industry's terminology, how you name your cases, what each abbreviation means.
Don't write the persona up front. Let the agent handle three or four real tasks, look at where it goes wrong, and add exactly that. A persona grown from observed mistakes beats any that was thought up in advance.
Having a persona suggested
You don't have to face the empty field. In the creation wizard's persona step the button reads “Suggest a persona”; in an existing agent's details it reads “Improve persona”. Both lead to the same flow.
- 1Optionally describe what the colleague should do for you — one sentence is enough. Without it, the suggestion is based on the name, role label and your templates; for an existing agent its knowledge base is taken into account too.
- 2Press “Suggest”. The suggestion appears as a card with a short rationale for why it is phrased that way.
- 3Read the text and edit it freely. Beside it you can expand the current persona and compare.
- 4“Apply persona” makes it the effective persona — for an existing agent as a new, traceable version; in the creation wizard the text is merely placed into the form, and creation still happens at the end. “Discard” leaves everything as it was.
The suggested text is readable by everyone in your organization once you apply it — the interface tells you so before you do. Don't put anything in a persona that not everyone on the team should see.
If someone else changed the persona in the meantime, or a newer suggestion superseded yours, bricksta does not let you overwrite blindly: you are asked to request a fresh suggestion so you decide on the current state. That is not a fault but protection against silently discarding a colleague's work. If the now-current state cannot be loaded at that moment, bricksta asks you to reload the page instead — the persona field is deliberately left untouched, so an empty field never looks like the saved state.
Suggesting also costs tokens from the agent's budget. On an agent with an empty budget, no suggestion is produced.
The knowledge base: context for all agents
Besides the persona, Agent Center → Agents holds a knowledge base: short, curated notes that your agents read and take into account when planning a mission. That way you give the whole team context once, instead of retyping it into every mission.
An entry has a title and a body and can be created with “New entry”, changed with “Edit” and removed with “Delete”. The checkbox “Active (agents read this)” controls whether an entry actually feeds in — a deactivated entry is kept but not read. The most recently updated active entries are taken into account first.
Keep the knowledge base short. Every active entry is read on every planning run and costs input tokens. Ten precise notes are better than fifty thorough ones — and they work more reliably.
The digital twin: an agent with a person's voice
An agent can be linked to a real person. It then becomes their digital twin and speaks from their material — with their vocabulary, their habits, their knowledge. The point of it is knowledge retention: experience that otherwise sits in one head becomes usable in cases.
A manager sets the link in the agent's details under “Digital twin of”: pick the linked person there and press “Save link”. From that moment, a “My twin” entry appears in that person's menu.
Only managers set the link — the linked person can neither link nor unlink themselves. They may maintain their twin afterwards, though: display name, role label, persona and the private knowledge. These changes too are versioned and visible to managers, who can override them if needed.
If the linked person has been deactivated or removed from the organization, the link persists and the private knowledge is retained — the interface says so. But take care: saving in that state would remove the link. Read the note before you confirm anything there.
Feeding the twin: taking in material
Under “My twin” — and for managers directly on the agent profile too — there is a “Take in material” section. What you feed in becomes private knowledge for that specific twin: only it draws on this, when planning, working and reviewing, in addition to the organization-wide knowledge base.
- Paste text
- Copy material straight in. Note in the field above where it came from — that helps you tidy up later.
- Upload files
- Text formats plus PDF and Word documents, up to 10 MB each. If a scanned PDF has no text layer, the AI reads it via image recognition; bricksta tells you explicitly so you check the text for recognition errors. If image recognition finds nothing either, the file is skipped and named.
- Import from bricksta
- Takes in the linked person's own bricksta comments with one click. If the agent is linked to nobody, this button is disabled — there would be no one whose comments could be taken in.
- Import from Outlook
- Takes in your own sent emails from the last 24 months, up to 200 of them. Sent only, never received, and your own mailbox only. Requires your Microsoft account to be connected under Settings → Connections.
- Distil
- Condenses the raw material taken in into a few reusable hints, replacing the raw entries. The path from “a lot of material” to “usable knowledge”. If nothing worth keeping is found, bricksta says so rather than inventing something.
Material taken in is explicitly presented to the AI as material, not as an instruction — a demand inside an uploaded file is meant to come to nothing. Don't rely on that alone, though: this boundary is a phrasing in the prompt, not a technical lock. The hard limits are and remain the tools, the approvals and the recipient selection from known contacts. So only take in material whose provenance you know.
Two things to know before uploading. First: the material is visible to people holding the “Manage organization” permission. Second, and this is the weightier point: it leaves bricksta. As soon as the twin works on a mission, its knowledge feeds into every step it carries out itself and thus goes to the AI provider — raw, never-distilled material included. On top of that come distilling itself, the persona suggestion, and image recognition for scanned PDFs. Don't upload anything that must not travel that route.
Your organization's own AI key does not cover all of these routes: missions, distilling and the persona suggestion accept it only from OpenAI and Anthropic, image recognition only from Anthropic and Google. Every other combination quietly runs through bricksta's central access. The section “The AI provider” in the article Setting up agents explains this in detail — check there before you take in material about individuals.
“Import from Outlook” appears only on your own twin under “My twin” — it imports your own mailbox, which only you can access. A manager cannot import someone else's mailbox, not even from the agent profile. Pasting text, uploading files and distilling, by contrast, are available in both places.
That limit does NOT apply to “Import from bricksta”: a manager can pull the linked person's bricksta comments from the agent profile without that person triggering it. Only the Outlook import is reserved to the linked person. If you have a works agreement or a data-protection policy, this distinction belongs in it.
If your Microsoft connection has expired or was never established, bricksta says so plainly and points you to Settings → Connections. If the import times out, try again right away; if older emails were not covered yet, the confirmation says so and you can run the import again to get further.
What an agent learns from its work
Independently of persona and knowledge base, agents gather experience: after each completed mission the lead agent draws brief lessons from how it went. In the agent profile you see them cumulatively under “What this agent has learned”, sorted by kind — worked, failed, template, delegation, cost or note.
These lessons arise by themselves and feed into future planning. You don't have to maintain them — but a look is worthwhile before you sharpen a persona: what an agent repeatedly records as failed usually belongs in its persona as a boundary.