Guía Completa de ChatGPT Agents

Main technological idea: ChatGPT Workspace Agents

The video frames ChatGPT Workspace agents as the next major upgrade after GPTs, aiming to replace them once fully released.

Unlike “GPTs” that mainly rely on provided context/instructions, agents behave like delegated AI workers that can:

• Plan an action workflow: produce a step-by-step “master plan” with checkmark completion.
• Run asynchronously: execute tasks while you’re not watching.
• Operate through multiple channels, not only the chat UI.

---

How agents differ from GPTs (product capabilities)

Agents are presented as having five core elements:

1. Goal (what “good” looks like for the task)
2. Trigger (one-off or scheduled/programmed reactions)
3. Instructions + Skills
   • process guidance
   • skills can enforce deterministic steps
4. Connectors / Applications (integration with internal/external tools)
5. Controls / permissions (fine-grained authorization, e.g., read Gmail but not send)

Key emphasized features

• Agents can use external applications and APIs.
• They connect through MCPs (tooling meant for agent integrations).
• Slack integration is supported so agents can work as part of a team workflow.
• Scheduling is supported (e.g., every 3 hours/daily) to collect info and deliver outputs—such as:
  • Slack reports
  • then creating tickets
• Memory inside the workspace agent configuration:
  • saves relevant interaction context over time
  • users can delete it.

---

Reviews / guides / tutorials covered

The video largely follows a step-by-step guide:

1. Where to find agents in OpenAI Workspace
   • Sidebar → “Agents” → “Explore agents”

2. Start from templates (examples include):

   • executive assistant
   • customer response writer
   • data analyst
   • sales assistant
   • SQL assistant (with advice to build custom agents for usefulness)

3. Create a new agent from scratch using natural language prompts

   • the agent builder auto-generates the plan, instructions, and tools
4. Edit agent instructions and add/update content
5. Add applications/connectors (e.g., Google Calendar/Drive/Gmail/Slack/Notion/GitHub/Teams)
6. Understand how skills work
   • deterministic processes
   • can be auto-generated by the agent builder
7. Upload and use reference files (e.g., PDFs) for evaluation
8. Manage permissions—especially dangerous actions (e.g., Gmail writing/sending/deleting)
   • described using controls like “always ask / never ask / custom”
9. Use testing vs preview
   • preview may produce more errors during beta
   • recommendation: test via normal deployment/workflow

---

Tutorial example #1: “Standards report reviewer” (document QA)

A concrete demonstration agent:

• Purpose: analyze an uploaded report against two company-standard PDFs:
  • a style/visual identity guide (typography, headings, formatting rules)
  • a content structure guide (required sections such as cover, control, abstract, executive summary, etc.)

Output behavior

• Returns a corrected/annotated document (“doc X”)
• Issues are marked in red, with detailed notes about contradictions.

Workflow detail

The agent follows a checklist-style process:

• review full report
• validate style
• validate structure
• create annotated file
• visually verify
• submit final + brief summary

Result claims (demo)

• Took ~2 minutes 20 seconds
• Detected specific missing or incorrect items, including:
  • confidentiality notice
  • table of contents rules
  • executive headline style contradictions
  • automated TOC level requirements
  • annex traceability requirements

Emphasis

The agent is described as preserving document formatting (e.g., Word documents) and acting like a “human editorial reviewer” based on the uploaded standards.

---

Tutorial example #2: “Drafts in Gmail” (customer support email automation)

Another demonstration agent:

• Connected to: Gmail
• Scheduling: runs periodically (e.g., every 3 hours)
• Function: scan inbox, determine which emails need responses, then generate draft replies for review.

Safety / permissions

• The video stresses configuring Gmail permissions to avoid risky actions.
• It recommends disabling or preventing dangerous permissions (e.g., deleting emails).

Demo behavior claims

• Drafts are created only for actionable emails.
• It skips “no-response-needed” automated confirmations (e.g., account opened on another computer).
• Took about 1 minute 44 seconds for the “task arrived” workflow.
• Drafts are tailored using historical response patterns from prior email data/training in the connected context (described as improving correctness).

Expansion path

• Could notify staff via Slack
• Then have humans send the emails.

---

Tutorial example #3: “AI Radar on X” (custom MCP + formatting skill)

A social listening agent for X (Twitter):

Why X integration is special

The video suggests X integration may not work well with generic internet search, so it uses:

• a custom MCP via an “XPI” app / developer platform approach

Integration approach

• Create a custom MCP through a workflow (video suggests using ChatGPT to implement MCPs)
• Cost estimates are mentioned:
  • ~€0.002 per query
  • €10 could cover years of queries (as claimed)

Output formatting

A skill (e.g., “Bulletlink Doc Formatter”) enforces a consistent structure:

• headline
• link
• explanation of why each item matters

Demo behavior

• Searches X for AI topics from the last two weeks
• Compiles results
• Outputs them in the required format

---

Iterative creation / quality control (agent builder loop)

The video highlights an optimization loop:

• Use agent builder to generate plan/instructions
• Upload documents and run tests
• If output is wrong, iterate:
  • adjust prompts/instructions
  • sometimes switch or generate skills

It also claims adaptive improvement: after repeated poor attempts, agent builder may recommend changing strategy—e.g., “create a skill” rather than only tweaking instructions.

---

Pricing / availability notes (important)

• Currently available for business accounts only (per the video).
• Agents are described as free until May 6th, then moving to paid token/credit-based usage.
• The video warns agents will be separate from subscription, and cost is not yet known.

---

Main speakers / sources (as identifiable)

• John (narrator/host; explicitly referenced during one demo)
• OpenAI (Workspace agents, connectors, MCPs, agent builder concepts)