Agent Orchestrator — bRRAIn Docs
Drag-and-drop AI orchestration that runs inside your bRRAIn — triggers, agents, decisions, sub-orchestrations.
Agent Orchestrator
A drag-and-drop visual builder for AI orchestrations that run inside your bRRAIn. Triggers, agents, decisions, sub-orchestrations, integrations across 28+ vendors, and full execution observability.
What it does
Agent Orchestrator lets non-engineers (and engineers) compose multi-step AI workflows on a canvas — pulling data from integrations, asking the brain Handler questions, branching on decisions, looping over collections, calling more orchestrations as sub-routines, and notifying humans when something needs review.
Common patterns built on it:
- Inbound trigger to action — a webhook arrives, the orchestration enriches the payload from your Vault, asks the Handler to classify, files a ticket in Linear or Jira.
- Scheduled assistant — every morning at 8 a.m., scan yesterday's commits, summarize for the team, post to Slack.
- Approval gate — a human approver reviews a Handler-proposed action before it commits.
- Multi-agent collaboration — three named agents with different roles cross-check each other before acting.
Installing
Free to install on every plan. Per-org subscription for high-volume usage above a generous free tier (10,000 runs / month). See the listing for current limits.
After installing, the extension appears in your Console under Installed extensions → Agent Orchestrator. The first time you open it, you're walked through a 3-step wizard:
- Pick a default LLM (the per-organization Handler is the default).
- Connect to one integration (Slack is the most common starter).
- Try a 2-node template (Manual trigger → Console Notification) to confirm everything fires.
The builder
The canvas opens with an empty grid. The left rail has the action library organized into seven categories:
- Logic — set variable, decision, delay, if/else, if/then, filter, fork, stop.
- Chain — sub-orchestration call.
- Notification — console, email, webhook.
- Functions — create file, vault append, HTTP request, send webhook, MCP call.
- AI — LLM, classify, extract, summarize, translate.
- Elements — sticky notes, group containers.
- Vault — read, write, query.
Drag any card onto the canvas to create a node. Each node has a slide-in config panel on the right edge of the canvas; click the node to open it, click the scrim to close.
Connect nodes by dragging from the right-edge handle of the source to the left-edge handle of the target. Edges support midpoint + insert: click the + to insert a new node between two existing ones.
Multi-select with drag-rect, then bulk-move or bulk-delete. Auto-layout button tidies the canvas. Linear-view button flattens the canvas into a step-by-step list.
Triggers
Each orchestration starts with a trigger:
- Manual — fire from the Run button.
- Scheduled — cron-like schedule with frequency / time / day-of-week pickers.
- Webhook — a unique URL we generate; signed payloads via HMAC.
- Vault event — fires when a record / file / decision / risk / project is added or changed (with path-glob narrowing).
- Integration event — fires on vendor webhooks (a Stripe payment, a GitHub PR, a Linear status change, etc.).
Agents
Agents are LLM-backed reasoning steps. Each agent has:
- A role (free-form description).
- A model chosen from the LLM Registry (the per-org Handler, an installed adapter, or a registered external model).
- An upstream-data picker that walks the canvas backward to surface available inputs.
- Free-form instructions in plain English.
Agents can be chained, can call sub-orchestrations, and can request human approval before their output commits.
Run history and observability
Every run writes a structured trace under extensions/agent-orchestrator/runs/<orch_id>/<run_id>.json:
- The trigger payload that started it.
- Each node's input and output.
- Errors with stack traces.
- The total duration.
- Any approval gates and their decisions.
The run history page lets you list runs, filter by status / time / trigger source, and replay a run with the same input.
Templates
The extension ships with 30+ starter templates organized by integration. Categories include:
- Slack integrations.
- GitHub workflows.
- Linear / Jira ticketing.
- Customer-support routing.
- Compliance evidence collection.
- Marketing automation.
Click Use template to load it onto the canvas. Customize the prompts, swap integrations, save as your own.
Sharing
Orchestrations can be shared:
- Within your organization (read or run).
- Publicly read-only (URL with the orchestration's name; gallery listings).
- Exported as JSON for import into another organization.
Scopes required
- Vault read and write in the
extensions/agent-orchestratorzone (auto-created on install). - Read access to any integration you wire into a card.
- Write access to any integration you wire into a card.
- LLM access via the per-organization Handler.
- Notifier write (for notification cards).
- Spawn one non-human actor named
agent-runner-bot.
Where to next
- Console: Marketplace install — managing this extension in the Console.
- Integrations — wiring the vendor connections used by orchestration cards.
- The extension's own in-app help (click
?from any builder page).