Video summary

Día 2 | Deja de hacer el trabajo tú y monta tu equipo de agentes de IA

Main summary

Key takeaways

Educational

Main ideas / concepts (Day 2: “Stop doing the work yourself—build your AI agent team”)

Purpose of the training (4-day AI Hero Live)

  • Move participants from basic AI use to professional, practical use.
  • Help them increase productivity and monetize AI knowledge.
  • Provide confidence they won’t be “left behind” by technology.

Key shift from Day 1 → today

  • Day 1: understanding AI, prompt engineering, and using tools like ChatGPT.
  • Today: the next step—AI agents (agentic systems) as the future of work.

Why agents matter (the problem they solve)

Companies accumulate “digital debt,” and work becomes broken. Common symptoms include:

  • Overstuffed inboxes
  • Too many meetings
  • Low engagement and fragmented communication (emails/Teams/Slack)

Human work splits into:

  • High-value work: analysis, research, design, strategy
  • Low-value coordination work: emails, meetings, chats

Goal: reduce coordination time so humans can focus on higher-value tasks.

Agents explained through analogies

Alexa vs. a humanoid robot (Optimus)

  • Alexa can answer, but doesn’t physically/fully execute tasks.
  • Agents represent the evolution that allows delegation from start to finish.

Agents vs. automation (e.g., n8n)

  • Automation is a fixed route from point A → point B.
    • If one step breaks, the whole flow fails.
  • Agents are adaptive:
    • they can reroute if something fails mid-process.

Agents as “virtual employees” (five ingredients)

An agent typically includes:

  1. Brain/model (GPT, Gemini, etc.)
  2. Knowledge (documentation/context provided)
  3. Tools (email, calendar, Slack/Teams, SharePoint, phone, etc.)
  4. Contract / instructions (system prompt: what to do and how)
  5. Triggers / autonomy (when they run; not just “you must type in chat”)

Agent networks remove bottlenecks (scaling work)

Typical bottleneck:

  • An overloaded employee dependency stalls others.

With agents:

  • Other agents produce drafts, research, and coordination quickly.
  • Work becomes parallel and scalable.
  • Agents can coordinate with each other—especially at higher levels.

“Platform” evolution analogy (mobile apps)

  • LLM models are like phone hardware.
  • Agents are like apps that make the underlying tech useful and productive.

Agent levels (classification)

Level 1 Agents (free / basic sandbox)

  • Where they live: inside an AI ecosystem (e.g., Gemini “Gems”)
  • Main characteristics:
    • Useful but limited
    • Often can’t freely connect to external tools
    • Focus: structured responses using role/context/instructions + uploaded documents

Level 2 Agents (tool-connected “productivity manager”)

  • Where they live: GPT ecosystem (custom GPTs / GPTs)
  • Main characteristics:
    • Can access external tools/applications
    • More autonomous and able to handle workflow items
    • Example tools: Gmail, Google Calendar, Google Drive, etc.

Level 3 Agents (advanced / multi-agent coordination + autonomy)

  • Main characteristics:
    • Coordinate with other agents (“an army”)
    • Iterate autonomously to a higher degree (as described)
    • Integrate across many apps to perform complex business workflows
    • Demonstrated using an external orchestrator platform (Relevance AI)

Methodology / step-by-step instructions taught

A) How to design a prompt for an agent (role template)

The speaker uses a structured prompt recipe:

  • Role
    • Define who the agent “is” (e.g., expert consultant, marketing agent)
  • Context
    • Define where it operates and for whom (team/company/market/user situation)
  • Task / Instruction
    • Break into stages/steps, such as:
      • Step 1: ensure understanding (ask follow-ups if unclear)
      • Step 2: consult knowledge/documentation; if missing, use internet (depending on settings)
      • Step 3: produce response and ask if more help is needed
  • Output format
    • Specify how answers must be returned (text, structured sections, document format, etc.)
  • Notes and examples
    • Add constraints about preferences and style (e.g., “summarize for Arnau,” “podcast style”)
    • Provide examples when beneficial
  • Constraints (most important safety/control)
    • Tell the agent what not to do, e.g.:
      • “Never make up information”
      • “Use only what exists in documents or verified sources”
      • “If a calendar slot doesn’t exist, don’t invent it”
      • “Don’t send emails without explicit approval/confirmation”

B) Level 1 build (Gemini “Gems” practice)

  • Access
    • Visit: gemini.google.com
  • Create a Gemini Gem
    • Go to Gems → Create / New Gem
  • Fill in agent configuration
    • Provide:
      • Name and description
      • Instructions using the role/context/steps/output/notes/examples/constraints structure
  • Upload knowledge
    • Add documentation files or a Google Drive folder
  • Save and test
    • Start a chat and ask:
      • “Who are you, what can you do, and what information do you have access to?”
  • Recommended testing query
    • Use a factual question that should be answered from documents (example shown: Tesla product range)

C) Level 2 build (ChatGPT custom GPT / GPTs with tools)

  • Access
    • Go to: chatgpt.com
  • Open GPTs / Explore GPTs
    • Click GPTs (described as an “App Store” for GPT agents)
  • Create
    • Click Create
    • Use an explicit instruction structure (role/context/steps/output/notes/examples/constraints), not only generic builder output
  • Configure
    • Set:
      • Name/description
      • Instruction prompt (role first, then context, steps, output format, constraints)
      • Output style (business-like, direct, structured)
      • Restrictions (especially “ask for confirmation before sending emails”)
  • Select model
    • Use the “thinking” model (example: GPT 5.5 thinking)
  • Connect tools/apps
    • Enable integrations such as:
      • Gmail (read/send as allowed)
      • Google Calendar
      • Google Drive
      • (optionally) other connected apps
  • Knowledge base
    • Upload PDFs/Excels/etc.
  • Run an end-to-end test
    • Example:
      • Check the calendar for the current week
      • Look up urgent emails
      • Check for documents in Drive
    • Validate that tool results return correctly

D) Level 3 demonstration (multi-agent orchestration with Relevance AI)

  • Concept
    • Use an orchestrator so:
      • One agent finds leads/clients
      • Another agent sends personalized proposals/emails
  • Flow example (Sanitas-like scenario)
    • User: “Find potential insurance clients in Valladolid…”
    • Agent 1
      • Uses tools (LinkedIn prospecting, web search, Google search, Excel/CRM creation)
      • Produces a list grouped by company/person
    • Agent 2
      • Receives leads from Agent 1
      • Sends individualized emails based on templates
  • Demonstrated outcome
    • Personalized subjects and tailored email content

Monetization / selling services (examples and workflow)

  • Big promise: agents can support income generation by enabling:
    • More output (productivity)
    • New service delivery (e.g., website redesign, agent-enabled services)

Level 3 monetization example: sales automation for a service

Using Cloud/Clot and agentic tooling:

  1. Find businesses with bad websites (via search)
  2. Create a redesigned website
  3. Generate a proposal (PDF)
  4. Send an email automatically after quote approval/confirmation

Workflow described in detail (website redesign loop)

  • Create a project in a Cloud design tool:
    • Search for company website URLs in a region
    • Build a small CRM
    • Prioritize prospects
    • Generate a Cloud Design prompt to recreate/upgrade the website
    • Export the webpage (PDF capture/budget)
    • Create pricing based on market ranges
    • Find contact email(s)
    • Send a commercial proposal email via an email connector

Speakers / sources featured (as mentioned in the subtitles)

Named speakers / presenters

  • Arnau (referred to as Arnau; co-host/leader; appears in VIP/Q&A mentions)
  • Javier (referred to as “Javi” and “Javier”; main presenter)
  • María de las Nieves Mérida Palomares
    • Story of using 40+ agents
  • Ricardo Rivera
    • Story of monetizing first AI agents (73 years old)
  • Ricardo’s story section narrator
    • Same person: Ricardo Rivera
  • Production / colleagues
    • Mentioned generally as “production colleagues,” no specific names

Other people/entities referenced

  • Ana María Goldenzal (submitted an example prompt)
  • Marina (submitted another example prompt)
  • Helmres / Hermes / Dior / Loe / Tesla / Porcelanosa / Learning Heroes / Sanitas / BUPA / Optimus (robot) / Alexa / Microsoft study
  • Relevance AI
  • Gemini / Google Gems
  • ChatGPT / GPTs / Custom GPTs
  • WhisperFlow (speech-to-prompt tool)
  • Clot / Cloud Design / Cloud Code
  • LinkedIn / Google / Google Drive / Gmail / Outlook / Teams / Slack / SharePoint / OpenTable
  • n8n (mentioned to contrast with agents)

Original video