Video summary
Aprende a construir agentes de IA para automatizar procesos.
Main summary
Key takeaways
Summary of key technological concepts & product features (Botmaker 3.0: “guidance agents”)
1) What Botmaker 3.0 is centered on: autonomous AI agents for process automation
Botmaker 3.0 is positioned as “100% agentic”, meaning it’s not a fully scripted chatbot/FAQ.
Instead, agents are designed to:
- Understand an objective and then guide a conversation to accomplish business processes
- Perform real actions in external systems while interacting with users
- Handle conversational flow that can branch, loop, go back, or pause, depending on missing information or context
2) Core architecture: Orchestrator + Channels + Agents
Botmaker 3.0 is structured around three main building blocks:
Channel
- The destination where the conversation starts/arrives (e.g., WhatsApp, webchat, social networks, email/voice).
Agent
- An AI-powered bot that completes a specific business process end-to-end.
- It can handle the full loop from collecting required information through execution and confirmation.
- Examples mentioned:
- Booking appointments
- Lead qualification, including level 1 support, shifts, and vacations
Orchestrator
An orchestrator agent acts like a director/conductor:
- Receives messages from the channel
- Selects the most relevant agents to consult (for token efficiency, it may consult only a subset)
- Synthesizes responses from one or multiple agents into a single reply
- Sends the final response back to the user—so output is centralized
Important behavior detail: The orchestrator is the one that speaks to the user. The agents primarily provide structured information to the orchestrator based on funnel steps/resources.
3) New/expanded vocabulary and how they map to building blocks
A glossary introduces key terms:
- Agent objective: What the agent must accomplish.
- Funnel: A step-by-step workflow (states/sequence) an agent follows to reach the objective.
- Example: ask for info → show availability → confirm → schedule → send reminder
- Not strictly “static”: it provides context, and the agent decides how to move through it.
- Logic: More detailed control inside funnels, described as ordered sequences that include:
- Instructions (e.g., “I need you to…”)
- Conditional blocks (conditions/actions expressed in natural language)
- Loops (repeat up to n times, mainly for iterative processing)
- Resources (tools the agent can use):
- Logical resources (re-usable “logic” pieces)
- Knowledge bases (e.g., PDF/Word documents for answering)
- MCP resources/integrations
- Existing Botmaker flows (older/static flows executed as resources)
- Multimodal: Agents can accept and respond using multiple input/output types:
- text, audio, images
- Automations: Process executions that can run:
- Periodically by time (e.g., every Monday)
- Triggered by conversation state (e.g., when the user confirms an order)
- May include non-message actions (e.g., updating a DB, saving to Google Sheets)
- MCP (Model Context Protocol):
- Connectors that let agents interact with external services/APIs.
- Native MCP examples mentioned:
- Google Calendar, email, Google Sheets
- Public MCP example mentioned:
- weather (to answer based on temperature/rain conditions)
- Data vs variables / memory approach:
- Variables aren’t required; the platform provides contextual memory across the conversation.
- If specific info is needed, it can be defined at funnel stages rather than stored as explicit variable memory.
- Human in the Loop:
- Inserts human approval/supervision at specific funnel states.
- The AI agent requests OK/rejection, the human validates, and then the AI passes the result to the user.
- The human doesn’t directly answer; the AI mediates for consistency.
4) Key differences from Botmaker 2.0 (as stated in the subtitles)
- Less rigid branching
- Previously: manually programmed decision branches/trees
- Now: define objective + funnel states + natural-language guidance; the flow is contextual, not a rigid decision tree
- Modular reusability
- Previously: flows/branches were harder to reuse across use cases
- Now: agents can be reused across channels/orchestrators, reducing rework
- Centralized tone/style
- In 3.0, tone and style are configured at the orchestrator level to avoid contradictions across multiple agents
5) Configuration and control features for orchestrators
The orchestrator can configure:
- Tone and personality (friendly, formal, playful, etc.)
- Language style (informal/direct, emoji usage constraints, etc.)
- Response length limits (e.g., max characters)
- Restrictions / rails: topics the agent should not cover (e.g., competitors, politics/religion, guarantee claims)
- UI steering: whether responses include buttons and can dynamically generate them
6) Example walkthrough shown in the platform UI (marketing agency scenario)
The walkthrough demonstrated Botmaker 3.0’s home screen:
- Old Botmaker 2.0 sections: bots with flows/mailbox/callbox
- New Guide agents: design orchestrators and agents
Example orchestrator:
- Connected to webchat
- Includes three agents for separate processes:
- Sales
- Frequently Asked Questions
- Customer follow-up
Sales agent (lead handling) funnel
- Step to request data
- Qualify lead using criteria and a scoring/rating approach
- If qualified: schedule a meeting/demo
- If not qualified: request email and send information
Funnel steps specify:
- required data fields
- who resolves each step:
- AI with instructions
- human supervision
- or handoff to human
Resources shown include:
- Knowledge bases
- MCP usage (e.g., email + calendar MCP to schedule meetings)
- Logic blocks and references to automations/external MCPs
7) Training path mentioned
- Presented as an introductory/theoretical session.
- Next session date mentioned:
- July 15: a more detailed tutorial on building agents to capture/qualify/convert leads using platform tools and integrations.
Main speakers/sources (from the subtitles)
- Seminar host/presenter (unnamed) speaking for Botmaker 3.0 / Guide agents
- Botmaker platform (referenced throughout as the product source)