On this page
Before you start
Only people with the “Manage organization” permission may create and manage agents. With “Manage members / roles” you can see the roster but change nothing — the page tells you so at the top with a “View only” note. If you cannot edit anything here, that is why.
- An agent needs an AI provider — which is satisfied out of the box by the central access bricksta provides. You only set up your own key if you want to (see below).
- An agent needs a budget greater than 0, otherwise it carries out no step at all.
- An agent needs at least the “AI” tool; everything else you grant deliberately.
- Your organization has a cap on how many agents can be active at once — ten by default.
Route 1: create a colleague from the starter fleet
Under Agent Center → Agents you find the “Starter fleet” section: eight prepared executor personas, ready on day one. Clicking “Create” turns one of them into an agent in your organization — with an identity, the tightly scoped executor role, matching tools and a starting budget of 500,000 tokens. A colleague that already exists is marked “Created” and is not made twice.
All eight receive the very same tightly scoped execution role. What distinguishes them is the persona — the instruction on how they work — not their permissions. After creation you can change anything on each of them: name, persona, tools, budget.
- Process lead
- Completes assigned tasks of a case promptly and keeps an eye on progress. Tools: AI, complete tasks. The obvious candidate when you later need a lead agent for missions — but you still have to grant it the “Run mission” tool explicitly.
- Process curator
- Works through recurring tasks while watching for patterns that could later become new templates. Tools: AI, complete tasks. Don't pick it to automate one specific specialist task — a narrower persona of your own is better for that.
- Account researcher
- Gathers customer and account context for onboarding handovers and answers research tasks in a structured way. Tools: AI, complete tasks. Not suitable if the research would need external sources — an agent works only with the context that lives in bricksta.
- CRM sync agent
- Works through tasks around keeping data aligned between bricksta and connected systems. Tools: AI, complete tasks. Important: it does not replace an integration — it works the tasks an integration produces. With no connection set up, it has nothing to do.
- Communications agent
- Drafts status updates and customer messages and sends them where granted. The only one in the fleet that additionally brings the “Send” tool (AI, complete tasks, send). Pick it only if messages really should go out — otherwise a colleague without send is the safer choice.
- Agent builder
- Helps with tasks around configuring further agents and templates. Tools: AI, complete tasks. It configures nothing itself — no agent may administer anything. It works the tasks that prepare such configuration.
- Knowledge-retention twin
- Preserves process know-how and answers tasks consistently from that knowledge. Tools: AI, complete tasks. The intended starting point for a digital twin — you set the link to a real person afterwards, on the agent itself.
- Platform concierge
- Answers questions about the application, explains features and helps with operating it. Tools: AI, complete tasks. Don't use it for substantive process work — its persona is aimed at explaining, not executing.
None of the eight receives the autonomy tools “Run mission”, “Work mission tasks” or “Send messages” at creation. That is deliberate: the boundary comes first, autonomy second. You grant those three later, individually, on the agent.
Route 2: build your own agent with the wizard
The “Create agent” button opens a wizard with a start screen and four numbered steps. You move through it with “Next” and “Back”; only the final button actually creates anything. Cancel before that and nothing comes into being.
- 1Start (not numbered): pick “Start blank” or “Start from a template”. Blank presets the “AI” tool and a budget of 500,000 tokens and leaves name and persona open. Starting from a template prefills name, role label, persona, tools and budget from one of the eight starter-fleet roles — you can change all of it afterwards. You cannot skip this screen; it is the fork in the road.
- 2Step 1 — Identity: display name (required, without it “Next” stays disabled) and role label (optional). The display name is what colleagues later see in the member list, on tasks and in the activity trail. Choose it so that the agent's remit is obvious.
- 3Step 2 — Persona: the core of the system prompt. Describe the task, the tone and the boundaries. Leave the field empty and the agent is still created — it then carries only the common behavioural frame that precedes every agent anyway, and works correspondingly unspecifically. You can refine the persona later at any time, and you can have one suggested here (see the article Personas and digital twins).
- 4Step 3 — Tools: exactly three are on offer here — AI, complete tasks, send. Only “AI” is preset. Pick none at all and you get a passive agent that can execute nothing; the wizard shows this as “No tools (passive)”. The three autonomy tools deliberately do not appear here — you grant those later on the finished agent.
- 5Step 4 — Budget and preview: the monthly token limit, preset to 500,000. Enter 0 and the wizard warns you immediately: budget 0 means the agent cannot carry out any task. Below it, a preview shows display name, role label, tools and budget — exactly what is about to be created. The “Create agent” button completes it.
Throughout, the wizard shows how many agent seats are used and free. If the cap is reached it says so plainly and the create button stays disabled — so you see it before you walk through the four steps for nothing.
Tools: what an agent may do at all
Tools are an agent's actual permission switch. They sit in each agent's details under “Tools” and are applied with “Save tools”. There are six, and they deliberately do not stack automatically — you grant each one individually.
- AI
- Allows the agent to use the AI model to think and draft. Without this tool it can do nothing at all. It is preset on every new agent.
- Complete tasks
- Allows it to actually complete tasks assigned to it. Without this tool an agent can think and propose, but cannot change the state of the case.
- Send
- Allows it to carry out a send task from your template. Whether a human approval is required is decided not by this tool but by the send task itself: in the template editor, “Trigger” is set either to “Confirm before sending” (the default) or to “Send automatically” — and with “Send automatically” the message goes out without approval. That choice exists only for the email route through bricksta; there is no automatic Outlook send. If you switch an existing step to Outlook, reset its trigger to “Confirm before sending” first — otherwise bricksta still picks the step up automatically and delivers it as an ordinary bricksta email, not through Outlook. Send proposals inside a mission sit behind an approval without exception anyway — that is what “Send messages” governs.
- Run mission
- Allows a lead agent to break a mission goal into a plan on its own. Without this tool you can create a mission but not start it — the picker marks such a lead as “not granted”.
- Work mission tasks
- Allows a participating agent to work through the mission tasks delegated to it. Without this tool its tasks stay open and the completion report lists them separately.
- Send messages
- Allows a participating agent to produce a send proposal within a mission work run. Deliberately separate from “Work mission tasks”, because sending has effects outside. Switching it on activates “Work mission tasks” automatically — without a work run there would be nothing to send.
An agent is never an administrator. There is no tool that grants administration — no roles, no members, no templates, no other agents. The persona changes nothing about that either: it governs HOW an agent works, never WHAT it may do. A text that awards itself permissions stays without effect.
Your organization's agent cap
Every organization has an upper limit on how many agents may be active at once. Without a specific arrangement it is ten. The wizard shows the state as “x of y active agents used — z free”. Once the limit is reached it shows instead that the cap has been hit, and creating is blocked — both in the wizard and for one-click creation from the starter fleet.
Two ways out: deactivate an existing agent — that frees a seat immediately, and its entire history is preserved. Or have the limit raised.
You cannot raise the cap yourself — not even with the “Manage organization” permission. That is built deliberately: otherwise it could be circumvented by deactivating, creating anew and reactivating. If you need more seats, talk to your contact at bricksta; the limit is then set for your organization.
The budget: how the entry comes into being
A budget is a monthly token allowance per agent. It comes into being two ways: the wizard creates it in the fourth step, and one-click creation from the starter fleet sets 500,000 tokens. Later you change it in the agent's details under “Budget (per agent)” via the “Token limit” field and “Save budget”. If it says “No per-agent budget set yet”, the agent has no entry of its own.
Budget 0 is not a pause button but a dead end: the agent still takes tasks on, every run is rejected by the budget gateway, and the task stalls as an incident. If you want to put an agent on ice temporarily, use the “Agent active” switch at the top of its details instead — that is the kill switch built for it.
How the budget behaves in operation — reserved versus consumed, what happens when it runs low, and how to recognize a rejected run — is covered in full in the article Budget, cost and runs.
The AI provider
For an agent to think, your organization needs an AI provider. Out of the box that is the centrally operated access bricksta provides — you need set nothing up for an agent to work. If you would rather use your own key, you store it under Settings, not in the Agent Center: it applies to the whole organization, not per agent. You enter the provider, the model and the key, and mark one of them as the default; usage then runs through your contract with that provider — though not on every route, see the note below.
Your own key is not used everywhere — and where it does not apply, the step quietly runs through bricksta's central access, with no error and no notice in the interface. Missions, the distilling of twin material and the persona suggestion accept it only from OpenAI and Anthropic. Image recognition of scanned PDFs only from Anthropic and Google. And the AI step in a template (the “AI” tool) always runs through central access — your key does not apply there at all. If you set it up precisely to control where data is processed, work through these routes individually.
Switch to your own key if your compliance rules dictate where data may be processed, or if you want to settle AI cost directly with the provider. For getting started, central access is the simpler route — you can switch later without touching the agents.
For the provider “OpenAI” there is an extra field, “Own endpoint (optional)”. It points bricksta at a self-hosted, OpenAI-compatible model — vLLM, Ollama, LM Studio and LiteLLM all speak the same interface. Leave it empty and api.openai.com stays in place. What is expected is the base address exactly as your server's own guide shows it (usually ending in “/v1”); whether you include the “/v1” makes no difference.
The address must be publicly reachable over https and must not carry a port. bricksta rejects internal targets — private networks, localhost and link-local addresses — before any request goes out at all. A model that runs only inside your company network is therefore deliberately out of reach: it needs a publicly reachable https entrance in front of it, typically a reverse proxy on 443. Use “Test connection” before you make the entry your default — if it fails, bricksta names the status code but not your server's response.
An agent's token budget counts tokens, not euros. It is a consumption cap, not an invoice — bricksta currently offers no conversion into cost. What a token costs at your provider is in their price list.
Setting up sending channels
If an agent is to prepare messages for the outside world, the channel needs setting up. Important to understand: the channel does not hang off the agent but off the send task in your template. Three routes are on offer there — email through bricksta, Teams, and Outlook. With nothing specified, it is email.
- The default route and the only one that works with no extra setup. Subject and body go out in full. Email is also the only channel for which bricksta writes back whether the message was actually delivered.
- Teams
- Reaches internal members only. The feed shows just the subject as a title plus a link to the case — the message body is not displayed there, it stays in bricksta. Requires an enablement for your organization (Microsoft tenant and catalog app stored, under Integrations). Unlike Outlook you can also pick “Send automatically” here: bricksta delivers Teams messages itself.
- Outlook
- Subject and body go out in full, but the prerequisites differ depending on who sends. If an agent sends during a mission run, the message comes from the mailbox of the person linked to it; that requires an enablement for the whole organization and fails without a link. If a person triggers the send task themselves on the case, it goes from their own mailbox — for which they only need their Microsoft account connected under Settings → Connections.
You enable Outlook yourself under Integrations: on the “Microsoft Outlook” card you enter your Microsoft organization's tenant ID and flip the switch. The permission required is “Manage organization”. Two things your Microsoft administration has to do FIRST, though: allow the bricksta app to send on behalf of your mailboxes, and confine that permission via an Exchange access policy to the approved sender mailboxes. Without that confinement the app would be allowed to send on behalf of every mailbox in your tenant. Your IT gets the technical steps from us.
How proven the channels are: the email route is in continuous use across bricksta and is well covered. Sending via Teams and via Outlook out of an agent run is built, but has barely run in practice — for Outlook we know of no successful send from the application. Try both with a harmless message to yourself before an agent reaches customers with them.
Teams and Outlook messages have no delivery feedback. There it stays at “sent” — bricksta cannot tell you whether the message arrived. Only for email does “Sent to …” later become “Delivered to …” or an error with a reason.
Done — and now what?
An agent with a persona, the tools “AI” and “Complete tasks”, and a budget above 0 is ready to work. Assign it a task in a case, the way you would assign one to a person, and see what happens. That is the best-proven route and the best way in.
Once that works, you can go further: the article Running missions shows how to hand a lead agent an entire goal. The article Budget, cost and runs explains how to keep an eye on operations — and what to do when an agent gets stuck.