Chapter 5: Multi-Agent Orchestration 🟡 BETA

Build complete applications through natural conversation. Describe what you want, and the system creates it automatically.

The AutoTask system uses an LLM-powered intent classifier to understand your request and route it to the appropriate handler. Whether you need a full web application, a simple reminder, or automated monitoring, you describe it in plain language.

AutoTask Architecture


Intent Types

TypeExampleWhat Gets Created
APP_CREATE“create app for clinic”HTMX pages, tools, schedulers
TODO“call John tomorrow”Task saved to tasks table
MONITOR“alert when IBM changes”ON CHANGE event handler
ACTION“email all customers”Executes immediately
SCHEDULE“daily 9am summary”SET SCHEDULE automation
GOAL“increase sales 20%”Autonomous LLM loop with metrics
TOOL“when I say X, do Y”Voice/chat command

Quick Start


Architecture Overview

Bot Database Architecture

One bot equals one database. All applications within a bot share the same data tables, tools, and schedulers.


File Structure

PathDescription
.gbdrive/apps/{name}/Generated web application
.gbdrive/apps/{name}/index.htmlMain HTMX page
.gbdrive/apps/{name}/assets/CSS, images
.gbdialog/tables.basDatabase schema definitions
.gbdialog/tools/Voice and chat commands
.gbdialog/schedulers/Timed automations
.gbdialog/events/Event triggers (ON CHANGE, ON EMAIL)

Creating an Application


Modifying Your Application

Use Designer to change anything about your app through conversation.


Adding Automation


Keywords Reference

KeywordPurpose
TABLEDefine data structure
FINDSearch records
SAVECreate record
UPDATEModify record
DELETERemove record
TALKSend message
HEARWait for input
SET SCHEDULECreate automation
ON CHANGEMonitor for changes
ON EMAILTrigger on email received

LLM / Chat Exposure (end-to-end reachability)

AutoTask is reachable from every channel (web, WhatsApp, Telegram) through the unified __api_call__ surface — the API command catalog. The LLM discovers the endpoints with api.find and executes them with api.exec; no BASIC keyword or per-bot prompt change is required.

Registered surface:

EndpointPurpose
POST /api/autotask/classifyClassify a plain-language intent into an automation plan
POST /api/autotask/compileCompile an automation plan into a BASIC script
POST /api/autotask/decideDecide immediate vs scheduled execution mode
POST /api/autotask/executeExecute a compiled automation plan
POST /api/autotask/create-and-executeOne-call automation: classify, plan, generate, execute
GET /api/autotask/tasksList AutoTask runs and their status
GET /api/autotask/statsRuns, success rate, pending approvals
POST /api/autotask/tasks/:task_id/approveApprove a pending task
POST /api/autotask/tasks/:task_id/cancelCancel a queued or running task

Curated chat commands (visible through apps.find/api.find):

  • tasks.autotask.create — turn plain language into a running automation (admin)
  • tasks.autotask.list — list runs and status
  • tasks.autotask.approve / tasks.autotask.cancel (admin)
  • tasks.autotask.stats

Security: mutating AutoTask endpoints are admin-only via the rbac_api_permissions matrix (/api/autotask% for POST/PUT/DELETE). The discovery filter (api.find) and the executor (api.exec) both enforce it, so regular users can list tasks but cannot generate or execute automation scripts. Read-only GET endpoints remain open.

Next Steps