What can an Agent do without asking?

Only what the Campaign's Autopilot level allows. The dial runs Off, Suggest Only, Auto with Approval and Full Auto, and it defaults to Suggest Only, so an Agent proposes and a person decides.

How do we stop an Agent that goes wrong?

With the kill switch, which halts one Agent or every Agent in the Workspace at once. Spending caps at run, day and month level stop a run before that, and a tool rule can deny a tool outright.

Every Agent action is written to the audit log with its reasoning, every run is stored step by step, and the Activity view replays a past run. Compliance reports export from the governance dashboard.

Does an Agent see personal data?

PII redaction is on by default, so personal data is scrubbed from logs and Agent inputs. Tool approval rules and the destructive action registry decide what an Agent may touch at all.

As role seats in the same Campaign roles a person can hold, not per token. The spending caps are governance controls you set, not a meter. The plans and their seats are on /pricing.

AI workforce and governance

Screenshot pendingThe Governance dashboard with the Spending Caps tab selected and caps configured at all three levels: the three cap levels with their values, and the tab row showing Spending Caps, Kill-Switch, Tool Allow/Deny Rules, Destructive Actions and Audit Log.

Agent Studio

An Agent is a named specialist with a role, a system prompt, an allowed tool set, a memory scope and a spending cap. Start from one of eight preset templates: Strategist, Copywriter, Designer Brief Writer, Media Buyer, Analyst, SEO Specialist, QA Reviewer and Community Manager. Every published configuration is an immutable version you can diff, roll back, publish, archive or duplicate.

Pick a preset, rewrite the prompt, and set the personality sliders.
Choose which tools the Agent may use, and attach knowledge bases with weighting.
Run the Agent against a scenario in the test panel and watch the step trace before publishing.
Set the per-run and per-day spending caps on the Agent itself.

Screenshot pendingAgent Studio on the create form with a preset template applied: the preset choice, the system prompt field, the tool picker, and the per-run and daily spending cap fields.

Tools

Everything an Agent can do is a declared tool. Built-in tools read Workspace data, write and schedule posts, search the Media Library, query connected analytics, read ad performance and publish to a content system. Teams add their own: a custom HTTP tool points at an internal API with domain allow-listing and credentials pulled from a secrets vault rather than pasted inline.

Upload an OpenAPI specification and let Collabre generate the tools from it.
Build a tool in the visual composer instead of writing configuration by hand.
Read tool analytics: which tools get called, how often, success rate and latency.

Screenshot pendingTool analytics for a Workspace with recorded tool calls, including one failure: the tool table with call counts, success rate and latency, and the failure detail drawer open on one row.

The human-in-the-loop Inbox

An Agent that needs something raises a blocking question, pauses its own run, and waits. There are six question types: Information, Approval, Choice, Creative Review, Tool Approval and Drift Alert. Questions route to the right person by expertise, carry a response deadline with a nudge at the halfway point, and expire unanswered.

Open the Inbox, read the question and the run it paused, and answer it.
Reassign a question, or act on several at once from the queue.
Answer with a saved template, private or shared, with variable substitution.
Record a spoken answer, then edit the transcript before you submit it.

Screenshot pendingThe Inbox on the Open tab with several questions of different types, one past its halfway nudge: the status tabs Open, Answered and Expired, the question type on each card, the Agent that asked, and the deadline.

Departments

A Department is a group of Agents with a charter, a conductor that decomposes the brief, and members that execute the subtasks. Seven presets ship: Marketing, Content, Analytics, Operations, SaaS Growth, DTC E-commerce and Agency. A person writes the brief, the conductor produces a plan, independent subtasks run in parallel, and the Department reports back with its cost and duration.

Write a brief, start a run, and watch subtask status on the run dashboard.
Swap a member Agent, or stack extra instructions onto one slot.
Clone a Department that worked, and read its performance scorecard.

Screenshot pendingA Department run in progress with subtasks in mixed states and one subtask paused on a blocking question: the conductor's plan, the subtask list with statuses, the paused subtask, and the run's aggregate cost and duration.

Activity

The Activity view shows what the workforce is doing right now. A real-time event feed carries filters and a step-trace drawer recording exactly what each Agent did. A kanban desks board groups the work into Department swimlanes, and a 3D office scene shows Agents at desks handing tasks over. Replay controls scrub back through a past run.

Filter the feed to one Agent and open its step trace.
Replay a run to the step where it went wrong.
Put two runs side by side and diff them.

Screenshot pendingActivity for a Workspace with several Agents mid-run: the live event feed, the filter row, and an open step-trace drawer on one event.

Memory

Memory has four scopes: session, Workspace, project and Department run. Workspace memory is opt-in and defaults to 30-day retention. There is deliberately no global memory across Workspaces, so nothing held in one Workspace reaches another. Cold-start snapshots let an Agent resume with context from its recent work rather than starting blank.

Set the memory scope and the retention when you configure the Agent.
Leave Workspace memory off, and the Agent keeps session context only.

Governance

Spending caps apply at three levels: per Agent run, per Agent per day, and per Workspace per month. Kill switches halt one Agent or every Agent at once. Tool approval rules allow, deny or require confirmation for any tool, scoped to the Workspace, an Agent, a Department or a Campaign, and the default is human confirmation on every call. Destructive actions are classified as automatic, human-in-the-loop or denied.

Set the Autopilot level per Campaign; it runs Off, Suggest Only, Auto with Approval and Full Auto, and defaults to Suggest Only.
Leave PII redaction on, so personal data is scrubbed from logs and Agent inputs.
Control which Workspace roles may edit which parts of an Agent's configuration.
Export a compliance report, and read each audit log entry with its reasoning.

Screenshot pendingThe Campaign Autopilot dial for a Campaign set to Suggest Only with a spending cap and an allowed-domain list: the four levels with the selected one marked, the per-Campaign cap, and the allowed-domain list.

Agent Marketplace

The Agent Marketplace lists Agent versions a Workspace has published, and Collabre seeds it with the same eight presets: Strategist, Copywriter, Designer Brief Writer, Media Buyer, Analyst, SEO Specialist, QA Reviewer and Community Manager. Installing clones the Agent into a Workspace you own or administer, so you can change its prompt, tools and caps without touching the original.

Browse the listings and read what each preset is set up to do.
Install a preset into your Workspace, then edit its prompt.
Publish one of your own Agent versions for other Workspaces to install.

Screenshot pendingThe Agent marketplace on the browse view with the eight seeded listings: the listing cards with their titles, and the install action on one card.

An AI workforce that works under your rules