Age of Ultron
autonomous AI does your desk work

139 / Two approaches to AI at work: casual use vs autonomous production. A live demo gives an AI agent browser access to Gmail, FreeScout, and a team page, then tasks it with re-engaging major customers over 20 years. It discovers hidden customer groups, identifies staffing agencies, and surfaces forgotten accounts—then a second agent joins the same browser.

Timeline

00:00 Two approaches to AI

01:14 What is Ultron?

01:38 IQ 70 meets IQ 200

01:53 Our configuration and setup

02:10 The business problem

02:59 The first prompt

03:43 Steering guidance while in-flight

04:04 Second query

04:19 Should you use a monster query?

04:35 The main task description

05:03 Prompt

05:50 Its first major insight

07:27 Using multiple browsers

07:40 Coaching and feedback

09:00 Resist the overlord!

09:32 Have the agent write its instructions

11:03 A trick for multiplexing agents

11:12 Multiple agents, one browser

Participants

@fulldecent

William Entriken

Episode notes

• GitHub Copilot — AI coding assistant used for the Ultron demo
• FreeScout — open source helpdesk where customer support interactions are stored
• VS Code — editor connected to Copilot for the autonomous agent workflow

Live demo of using autonomous AI agents (“Ultron”) for real business work at a medical training company.

• Two approaches to AI in business: “working” (interactive, ad hoc) and “operational” (autonomous, production-grade)
• The “Ultron” pattern: give an AI agent access to your browser and let it operate across Gmail, FreeScout, team pages, and Google Docs
• Analogy: you are someone with an IQ of 70 talking to someone with an IQ of 200 — frame your prompts accordingly, keep them simple, and let the model figure out the details
• Business task: re-engage the largest customers over the past 20 years
  • Build a list from a team page
  • Cross-reference with Gmail and FreeScout for conversation history
  • Identify customer groups, staffing agencies, and duplicate accounts
  • Recognize the difference between group administrators buying for others versus individuals buying for themselves
  • Spot customers who changed email domains
• Steering the agent mid-task: give it corrections and hints while it works, rather than one huge upfront prompt
• The agent autonomously discovered business insights: individual expiration cycles vs group-level purchasing, hidden staffing agency patterns, and forgotten customer accounts
• Having the agent write its own instruction file: tell it to document everything it learned so another agent (or person) could pick up the project cold
• Running multiple Ultron instances in the same browser session:
  • Each instance tags its own tabs using a URL fragment or query string parameter like ?ultron-instance=2
  • Avoids the overhead of creating separate browser profiles and re-authenticating everywhere
• “That is table stakes for 2026”