Keynote: Will the Real Autonomous Agent Please Stand Up
summary
Argues that dependable AI is a validation problem, not raw capability, and shows Microsoft shipping Scout autopilot and copilots as examples that run in the background and reshuffle work. Recommends running experiments (PI: People, Ingredients, Experiments), suggests 25 experiments over a season, and estimates five runs before it works. Names concrete new roles—testing agent boss, guardrail guru, context captain—and warns token-maxing projects died once cloud bills arrived.
AI-written summary · about
chapters (40)
- 0:00Start
- 0:04Keynote Opens - Will the Real Autonomous Agent Please Stand Up?
- 0:34Framing: Dependable AI Is a Validation Problem, Not a Capability One
- 1:06Speaker Intro - Dona Sarkar, Chief Troublemaker, Microsoft
- 2:08Why What Worked Last Year Won't Work This Year
- 3:41The Guardian Headline About the Robot Apocalypse - From 2019
- 4:41The "We Need to Do AI" Phase Every Company Went Through
- 5:12Token Maxing - and the Summer It Died When the Bills Arrived
- 5:43Faster Horses: The Quote Henry Ford Never Actually Said
- 6:45How Operating Systems Used to Ship: 36-Hour Builds and Mailed Discs
- 7:48Faster Horses Optimise the Present; Cars Create a Future
- 8:49The First iPhone Apps Had Scroll Bars
- 9:20Every Era Has Its Scroll Bar Moment
- 9:51Why AI Is Hard to Reimagine: There Is No New Form Factor
- 10:22What Broke the Scroll Bar: Shazam and the Big Blue Button
- 11:52Why the Replacement Framing Is Wrong
- 12:23If Your AI Project Works as a 2016 Startup Pitch, It's a Faster Horse
- 12:53The Assumptions AI Lets Us Make
- 13:23Form Factors Testers Will Have to Cover
- 13:53New Category One: Software With No UI, Built for Agents
- 14:25New Category Two: An App That Forks a Billion Different Ways
- 14:55New Category Three: Mortal Software That Spawns Agents and Vanishes
- 15:56How to Use AI Usefully: The Spell Check Analogy
- 16:57The Three Phases: Chat, Artifact, Auto
- 17:27Copilots Are Reactive; Autopilots Are Proactive
- 18:31How Do You Test That? We Don't Know
- 19:02The Fashion Analogy: Off the Rack, Tailored, Couture
- 21:04If You're Doing Work AI Can Do, You're Competing With Yourself
- 21:35Are All the Jobs Going Away? OpenAI Just Hired a Head of Recruiting
- 22:36The PI Framework: People, Ingredients, Experiments
- 23:37The Question: What Couldn't We Solve Last Year That We Can Now?
- 24:38Real New Jobs: The Testing Agent Boss
- 25:38The Guardrail Guru, the Context Captain, the Taste Titan
- 26:08The Intent Translator - Because Nonsense In Means Nonsense Out
- 26:39Human-Created Will Carry a Premium; Human-Curated Will Be Valuable
- 27:12Homework Solo: List Your Work and Split It Three Ways
- 27:42Homework for Managers: Run 25 Experiments Over One Season
- 28:44Q&A: Minimising Hallucination Cascades in Autonomous Agents
- 29:14Three Defences: Clear Instructions, AuthN/AuthZ, Log Everything
- 30:47Expect Five Runs Before It Works - Think Video Game, Not Linear Software
related talks
01h 01mKeynote Panel: Future of AI agents
01h 17mThe year of the AI Agent: Beyond Automation
01h 00mKenote: AI Didn’t Replace You. It Promoted You!
19mKeynote: Towards Building Safe and Secure Agentic AI
54mKeynote — Noise and Signal
27mHow to Build AI Agents and Keep Your Sanity
29mPairing with the machine: How AI Agents are rewriting the rules of software development
24mKeynote - The AI Playbook For Companies