On this page
The Agent Center is young. It is fully built, but it has only been used in a handful of organizations and across a small number of runs. Start small: one agent, one clearly bounded task, a modest budget — and study those first runs under “Runs & Cost” closely before you hand over more. Each of the following articles flags the functions with the least real-world experience behind them.
What an AI agent is — and isn't
An agent is a member of type AI agent: it appears in the member list with a robot icon and shows up in the activity trail as an actor — traceably marked “(AI agent)”, so it is always clear whether a human or an AI acted. It works on tasks assigned to it within your processes, rather than acting freely “somewhere”.
An agent gets a tightly scoped executor role: start, fill in and complete tasks — nothing more. It cannot manage anything, configure anything or change other members. And it only acts on tasks it is explicitly responsible for; an agent does not automatically grab “ownerless” tasks. That keeps its reach narrow even when several agents work in the same organization.
An AI agent cannot sign in interactively. There is no agent login anyone could misuse — agents act solely through the controlled execution path inside bricksta.
As with the manual way, it is not the agent that decides whether a step needs a human approval — you do. In a template, low-risk steps such as filling in fields run straight through, and for a send step the send task itself decides whether it must be confirmed before sending (the default) or not. In a mission you decide per initiative whether field suggestions come to you first — reviewing is the default — and every send proposal there sits behind your approval without exception.
The model behind it: you are the orchestrator
The picture underlying the whole surface is a small workforce: you set the direction, the agents carry it out. Concretely that plays out across four levels, each building on the last.
- 1A single task: you assign an agent a task in a case, just like a person. It fills the task in and completes it. This is the simplest and best-proven path.
- 2A mission: you state a goal rather than a task. A lead agent breaks it into a plan you approve, and hands the tasks to a team of agents.
- 3Autonomous execution: the participating agents work the delegated tasks themselves — field values, notes, message drafts — each only with the tools you granted them individually.
- 4Keep working on its own: a schedule continues a running mission even when nobody has the app open. The plan, the completion and every send proposal remain yours to decide.
Every level is its own deliberate switch. A new agent can initially only think (the “AI” tool) — not complete, not send, not run a mission. You grant each further ability on purpose, per agent. There is no path by which an agent grants itself a permission.
The words you need
- Agent (agent profile)
- A named AI coworker: display name, role label, persona, tools, budget and an active/inactive switch. Lives under Agent Center → Agents.
- Persona
- The core of the agent's instructions — it governs HOW it works (task, tone, boundaries). It never governs WHAT it may do; that is settled by tools and budget alone. Persona changes are versioned.
- Tool
- An individually grantable ability: AI, complete tasks, send, run mission, work mission tasks, send messages. Without the matching tool the run stops before anything happens.
- Budget
- A monthly token allowance per agent. It is checked before every run and reserved for the duration of that run. If it doesn't cover the estimate, the step never starts.
- Run
- A single execution attempt by an agent on a task, with its outcome and the tokens it consumed. Every run is visible under Agent Center → Runs & Cost — including the denied ones.
- Mission
- An initiative with a goal instead of a task list. It has a lead agent, participating agents, optionally a linked case, and two approvals by you: the plan up front and the completion at the end.
- Approval
- An agent action waiting for your decision. There are two kinds: action approvals (under Agent Center → Approvals, e.g. an email from a template action) and mission approvals (on the mission detail: plan, completion, field suggestions, send proposals).
- Digital twin
- An agent linked to a real person that speaks from that person's own material. The linked person maintains it themselves under “My twin”.
- Keep working on its own
- A switch per mission. With it on, a schedule picks the mission up regularly and carries it on. It never starts a mission, approves neither the plan nor the completion, and never sends any of the messages your team drafts. What runs alongside that out of your template is covered in the article Keep working on its own.
The four views of the Agent Center
You find the Agent Center under “Work” in the menu. It has four views; which ones you see depends on your permissions. Opening the center without a sub-view lands you on “Approvals”.
- Agents
- The roster of your AI workforce: create, persona, tools, budget, knowledge base, gathered experience, twin link. Visible with the “Manage members / roles” or “Manage organization” permission; changing anything requires “Manage organization”.
- Approvals
- Every agent action waiting for a human decision, in one place. Visible to all staff — each person sees only what they are allowed to decide.
- Missions
- Creating and running missions. Requires the “Manage missions” permission.
- Runs & Cost
- Activity and token consumption per agent, the remaining budget, the organization pool, and the log of missions that ran on their own. Requires the “Manage organization” permission.
Agents are a special member type with a fixed executor role. How that fits into the permission model (types → roles → rights) is explained in the article Permissions and visibility.
Where to go next
The remaining articles in this chapter are built as one continuous path — from nothing to a mission that keeps running on its own. Read them in order if you are starting out, or jump straight in.
- 1Setting up agents — the creation wizard step by step, the eight roles of the starter fleet, tools, your organization's agent cap, the AI provider and the sending channels.
- 2Budget, cost and runs — how a budget comes into being, what reserved and consumed mean, what happens at budget 0, and how to read the “Runs & Cost” table column by column.
- 3Running missions — state a goal, approve the plan, delegate, review field and send proposals, complete. The longest article, because this is where the actual work happens.
- 4Personas and digital twins — have a persona suggested, link an agent to a real person, feed in material and distil it.
- 5Keep working on its own — the schedule behind a mission: what it does, what it explicitly does not do, the five pause reasons and the way out of each.
Start with a single, clearly bounded task. An agent that reliably handles exactly one recurring activity is worth more than a broadly authorized one nobody quite trusts.